Shiki42 commited on
Commit
a174e5c
·
verified ·
1 Parent(s): 7b4c0ee

Publish E002 SE(3) spline recovery experiment

Browse files

Protocol-frozen E002 baselines and validation-selected minimum-jerk checkpoint; fresh-test conclusion rejected.

Files changed (42) hide show
  1. README.md +8 -8
  2. SHA256SUMS.json +203 -92
  3. experiments/E002/E002-CARD.md +58 -0
  4. experiments/E002/audits/action-target-audit.json +64 -0
  5. experiments/E002/audits/data-split-audit.json +191 -0
  6. experiments/E002/audits/proposal-manifest.json +0 -0
  7. experiments/E002/conditions/c-endpoint-5d2dff166d/CONDITION_CARD.md +15 -0
  8. experiments/E002/conditions/c-endpoint-5d2dff166d/best.pt +3 -0
  9. experiments/E002/conditions/c-endpoint-5d2dff166d/config_resolved.yaml +111 -0
  10. experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_fresh_test.json +278 -0
  11. experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_validation.json +278 -0
  12. experiments/E002/conditions/c-endpoint-5d2dff166d/run_manifest.yaml +93 -0
  13. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/CONDITION_CARD.md +15 -0
  14. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/best.pt +3 -0
  15. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/config_resolved.yaml +114 -0
  16. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_fresh_test.json +295 -0
  17. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_validation.json +295 -0
  18. experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/run_manifest.yaml +104 -0
  19. experiments/E002/conditions/c-residual-h16-f186e90df2/CONDITION_CARD.md +15 -0
  20. experiments/E002/conditions/c-residual-h16-f186e90df2/best.pt +3 -0
  21. experiments/E002/conditions/c-residual-h16-f186e90df2/config_resolved.yaml +114 -0
  22. experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_fresh_test.json +295 -0
  23. experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_validation.json +295 -0
  24. experiments/E002/conditions/c-residual-h16-f186e90df2/run_manifest.yaml +104 -0
  25. experiments/E002/experiment-registry.yaml +214 -0
  26. experiments/E002/experiment-summary.json +914 -0
  27. experiments/E002/fresh-test-firewall-receipt.json +30 -0
  28. experiments/E002/fresh-test-summary.json +290 -0
  29. experiments/E002/protocol-freeze-receipt.json +22 -0
  30. experiments/E002/runs/E002-R001/condition_config.yaml +109 -0
  31. experiments/E002/runs/E002-R001/evaluation_validation.json +278 -0
  32. experiments/E002/runs/E002-R001/run_manifest.yaml +93 -0
  33. experiments/E002/runs/E002-R002/condition_config.yaml +112 -0
  34. experiments/E002/runs/E002-R002/evaluation_validation.json +295 -0
  35. experiments/E002/runs/E002-R002/run_manifest.yaml +104 -0
  36. experiments/E002/runs/E002-R003/condition_config.yaml +112 -0
  37. experiments/E002/runs/E002-R003/evaluation_validation.json +295 -0
  38. experiments/E002/runs/E002-R003/run_manifest.yaml +104 -0
  39. experiments/E002/runs/E002-R004/condition_config.yaml +112 -0
  40. experiments/E002/runs/E002-R004/evaluation_validation.json +295 -0
  41. experiments/E002/runs/E002-R004/run_manifest.yaml +104 -0
  42. experiments/E002/selection-validation.json +109 -0
README.md CHANGED
@@ -5,18 +5,18 @@ tags:
5
  - robotics
6
  - action-chunking
7
  - se3
 
 
8
  - robotwin
9
  datasets:
10
  - Shiki42/up-vla-precision-recovery-15k
11
  ---
12
 
13
- # UP-VLA Action Chunk Recovery Checkpoints
14
 
15
- Protocol-registered E001 artifacts for endpoint, direct H16, and residual H16 recovery.
16
- The selected residual model improved grouped test path score by 4.68%, below the
17
- pre-registered 10% validation gate; E002 was not opened.
18
 
19
- The source dataset includes RLBench-derived material with non-commercial research
20
- restrictions. Treat these checkpoints as research-only unless all upstream license
21
- obligations have been independently resolved. See experiments/E001/E001-CARD.md
22
- and the immutable run/evaluation receipts for details.
 
5
  - robotics
6
  - action-chunking
7
  - se3
8
+ - b-spline
9
+ - minimum-jerk
10
  - robotwin
11
  datasets:
12
  - Shiki42/up-vla-precision-recovery-15k
13
  ---
14
 
15
+ # UP-VLA Structured Recovery Checkpoints
16
 
17
+ Protocol-registered recovery experiments:
 
 
18
 
19
+ - E001: endpoint/direct/residual H16 action chunks; completed inconclusive.
20
+ - E002: same-split endpoint and residual H16 baselines plus clamped cubic and true quintic minimum-jerk SE(3) residual heads; completed rejected. The selected minimum-jerk head reduced fresh-test grouped jerk by 18.50% but missed the frozen 25% effect gate and worsened path score by 4.18%, above the 2% bound. Real Pi0.5 rollout is not authorized.
21
+
22
+ The source dataset includes RLBench-derived material with non-commercial research restrictions. Treat these checkpoints as research-only unless all upstream license obligations have been independently resolved. See each experiment card and immutable run/evaluation receipts.
SHA256SUMS.json CHANGED
@@ -1,103 +1,214 @@
1
  {
2
- "experiment_id": "E001",
3
- "file_count": 24,
4
- "files": {
5
- "README.md": {
6
- "bytes": 719,
7
- "sha256": "929577f73e70f2a9404890a64881e3aefb423dd06f8cba19d946957a9222c46b"
8
- },
9
- "experiments/E001/E001-CARD.md": {
10
- "bytes": 1516,
11
- "sha256": "0ee76cd69ac690653483f469f55ebaf16c3ebcc5c9f841a480909bed2ae87e6e"
12
- },
13
- "experiments/E001/audits/action-target-audit.json": {
 
 
 
14
  "bytes": 1741,
 
15
  "sha256": "16e1b62f9f54fb0ac8a0fab0b7772e66ab6b6bc0d930fc5aa30f53ff19258fa6"
16
  },
17
- "experiments/E001/audits/data-split-audit.json": {
18
- "bytes": 1360,
19
- "sha256": "db98be16cd36ccb1726d5407202b126f8ed147e31144ba6c1228dc163e8be353"
 
20
  },
21
- "experiments/E001/audits/real-data-smoke.json": {
22
- "bytes": 934,
23
- "sha256": "98502ae238c997ed151d99bbf8caf1c0e1bfad04be181f422deef73fc8fa4505"
 
24
  },
25
- "experiments/E001/audits/residual-zero-init-smoke.json": {
26
- "bytes": 671,
27
- "sha256": "6aa37549464368866c756979073f38f79b45cf567462e19b8155eba5fa255c30"
 
28
  },
29
- "experiments/E001/conditions/c-endpoint-3c2dae9044/CONDITION_CARD.md": {
30
- "bytes": 559,
31
- "sha256": "4b2ef7a1cb78c451600706169a40d2deff071264cf5fee1374c2dd2a2c34d7a2"
32
- },
33
- "experiments/E001/conditions/c-endpoint-3c2dae9044/best.pt": {
34
  "bytes": 189704503,
35
- "sha256": "182fc8544a1489582c7f96e29625b7e0048642edf6de749c50b19da0da33379f"
36
- },
37
- "experiments/E001/conditions/c-endpoint-3c2dae9044/config_resolved.yaml": {
38
- "bytes": 2741,
39
- "sha256": "a2626ca28030873c0ce82e3f9a2c6057e976f3f89d9c693c4c48561b02183b9e"
40
- },
41
- "experiments/E001/conditions/c-endpoint-3c2dae9044/evaluation_test.json": {
42
- "bytes": 3871,
43
- "sha256": "4bb067239611e3f9427ba5fffdc8c0b56a30452f6db8e0a02b2ade56a2b63e43"
44
- },
45
- "experiments/E001/conditions/c-endpoint-3c2dae9044/evaluation_validation.json": {
46
- "bytes": 3879,
47
- "sha256": "2733fe8093af6a9fad8f9c6988caee1cd048412c3620ad193779a27bcdbe4972"
48
- },
49
- "experiments/E001/conditions/c-endpoint-3c2dae9044/run_manifest.yaml": {
50
- "bytes": 3472,
51
- "sha256": "6548cd5b4b3335d5b5f88262924e34a8fcbf0621216792bc946bad804026c418"
52
- },
53
- "experiments/E001/conditions/c-residual-h16-be2e741500/CONDITION_CARD.md": {
54
- "bytes": 567,
55
- "sha256": "c244e4d9a3b123628c29e3bba41e378a38f48ad6d24fe60fb62bba7d86b35792"
56
- },
57
- "experiments/E001/conditions/c-residual-h16-be2e741500/best.pt": {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  "bytes": 238782079,
59
- "sha256": "81ef4841165e996bf7c94556e3f83b457ae5b021bfa6c99ae2fa4998e3f111f3"
60
- },
61
- "experiments/E001/conditions/c-residual-h16-be2e741500/config_resolved.yaml": {
62
- "bytes": 2955,
63
- "sha256": "36d1731639d935f7cd94d6a01c67205577985242498517a6bcf046f78ce0bdde"
64
- },
65
- "experiments/E001/conditions/c-residual-h16-be2e741500/evaluation_test.json": {
66
- "bytes": 4660,
67
- "sha256": "7f77315929e49166cf976cc00d3f8de8b157a35a7bdd17a507763bca2d6bc813"
68
- },
69
- "experiments/E001/conditions/c-residual-h16-be2e741500/evaluation_validation.json": {
70
- "bytes": 4657,
71
- "sha256": "9373e8bd953247345671b604f00945aa5649b2a8887a94136547e3de77f6a0d4"
72
- },
73
- "experiments/E001/conditions/c-residual-h16-be2e741500/run_manifest.yaml": {
74
- "bytes": 4186,
75
- "sha256": "c97c0ce3f3ebfeccce4c75b1f9d9a039f462fa2812b252169a27e8f694667fe7"
76
- },
77
- "experiments/E001/experiment-summary.json": {
78
- "bytes": 16356,
79
- "sha256": "bc3ef040d1f2d9e65347a56c1079e027dbde9c6aaee39ea48b10f1503878aa7d"
80
- },
81
- "experiments/E001/runs/E001-R002/evaluation_validation.json": {
82
- "bytes": 4349,
83
- "sha256": "e69679a21e423a975480369d3c9773a812bff06298f81ab6a26e2004fe1bc338"
84
- },
85
- "experiments/E001/runs/E001-R002/run_manifest.yaml": {
86
- "bytes": 4001,
87
- "sha256": "ba17d383fee1444110197648235d00d447ef14055df8f7ab9280c84486316fd7"
88
- },
89
- "experiments/E001/runs/E001-R003-invalid/metrics.jsonl": {
90
- "bytes": 195729,
91
- "sha256": "7cbed4f6a4fe6871e0ea2ecadeec972e56775ef1220ea1903c7beb672ede8a96"
92
- },
93
- "experiments/E001/runs/E001-R003-invalid/run_manifest.yaml": {
94
- "bytes": 4259,
95
- "sha256": "b27935c72939dc5112d422ab19c9b422b539bd041cc4ad6f3c2f5546001b24e8"
96
- },
97
- "experiments/E001/selection-validation.json": {
98
- "bytes": 2021,
99
- "sha256": "fc742f723daf1f6e91b78db104c78c47056ad965fc0510de8ea9fe4964d868b3"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100
  }
101
- },
102
- "schema_version": "upvla.checkpoint_package.v1"
 
103
  }
 
1
  {
2
+ "content_bytes": 669255513,
3
+ "experiment_id": "E002",
4
+ "file_count": 41,
5
+ "files": [
6
+ {
7
+ "bytes": 960,
8
+ "path": "README.md",
9
+ "sha256": "42eb4ce99074d508a9e5551ff4f7f74a8c8488f29b6a164ac623defc5a5f123c"
10
+ },
11
+ {
12
+ "bytes": 4564,
13
+ "path": "experiments/E002/E002-CARD.md",
14
+ "sha256": "03bca7ca8ffef76374670c60f5884dc3ee2f5a576023bfb8106a501046093c6c"
15
+ },
16
+ {
17
  "bytes": 1741,
18
+ "path": "experiments/E002/audits/action-target-audit.json",
19
  "sha256": "16e1b62f9f54fb0ac8a0fab0b7772e66ab6b6bc0d930fc5aa30f53ff19258fa6"
20
  },
21
+ {
22
+ "bytes": 10481,
23
+ "path": "experiments/E002/audits/data-split-audit.json",
24
+ "sha256": "022dac36e48ce65ea920a789a36e1458b40e78df7f55eabe15504fd8ef0f96be"
25
  },
26
+ {
27
+ "bytes": 1782289,
28
+ "path": "experiments/E002/audits/proposal-manifest.json",
29
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
30
  },
31
+ {
32
+ "bytes": 958,
33
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/CONDITION_CARD.md",
34
+ "sha256": "a708300ca71964185012cb07e3cd8052d7207cb43129b4be96101b7397d68f2a"
35
  },
36
+ {
 
 
 
 
37
  "bytes": 189704503,
38
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/best.pt",
39
+ "sha256": "0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da"
40
+ },
41
+ {
42
+ "bytes": 3281,
43
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/config_resolved.yaml",
44
+ "sha256": "9e752d58616638c78f1664d4379202ad607db97c8bdafb71671a0634e7441042"
45
+ },
46
+ {
47
+ "bytes": 10895,
48
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_fresh_test.json",
49
+ "sha256": "6bac903b2a7a547f753130a77604bafda1e4f76a8138396734c0bd98bd7187df"
50
+ },
51
+ {
52
+ "bytes": 10903,
53
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_validation.json",
54
+ "sha256": "0fda3f48a9f3ac61cdc91d4e13c0cd0d7bbb006c23a4f074d0426313da506b06"
55
+ },
56
+ {
57
+ "bytes": 3518,
58
+ "path": "experiments/E002/conditions/c-endpoint-5d2dff166d/run_manifest.yaml",
59
+ "sha256": "d273fe41b8dfcbf020c4048aeec3e47296de4bf13bc56bef369df56a8d6c328b"
60
+ },
61
+ {
62
+ "bytes": 1006,
63
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/CONDITION_CARD.md",
64
+ "sha256": "1847112ee7e24a1d8a94f1d055c5531c5ecb8e8b04f0eb07b5d20394a09296a7"
65
+ },
66
+ {
67
+ "bytes": 238735999,
68
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/best.pt",
69
+ "sha256": "4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9"
70
+ },
71
+ {
72
+ "bytes": 3481,
73
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/config_resolved.yaml",
74
+ "sha256": "adcecfd2a30622bbc900b3b1397035eb7ce20bf94c8fce83b67aade2fa16e5d8"
75
+ },
76
+ {
77
+ "bytes": 12018,
78
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_fresh_test.json",
79
+ "sha256": "8c32ee98a9039c7a5f426f44c5aec1eec00d65e06347a15d2ed9899387d2e4c7"
80
+ },
81
+ {
82
+ "bytes": 12025,
83
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_validation.json",
84
+ "sha256": "7c451bf90cb7e2f1ea50a3dff1647bca8a9cea6d4adf9003733d8140e6beb741"
85
+ },
86
+ {
87
+ "bytes": 4247,
88
+ "path": "experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/run_manifest.yaml",
89
+ "sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac"
90
+ },
91
+ {
92
+ "bytes": 903,
93
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/CONDITION_CARD.md",
94
+ "sha256": "b68b390848269884198b932df495579acdc273e647981e6353f2ab1398969d16"
95
+ },
96
+ {
97
  "bytes": 238782079,
98
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/best.pt",
99
+ "sha256": "c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9"
100
+ },
101
+ {
102
+ "bytes": 3467,
103
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/config_resolved.yaml",
104
+ "sha256": "6413c33b4a7db008f8c01f024bfc62e8d191a7240f321913a720852804116fd6"
105
+ },
106
+ {
107
+ "bytes": 12009,
108
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_fresh_test.json",
109
+ "sha256": "390e69b77677e4e6096afa2eefde8edb633f88558746f715ecbfb9ed03102742"
110
+ },
111
+ {
112
+ "bytes": 12007,
113
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_validation.json",
114
+ "sha256": "17a116944f1967d60d2427975690f635f4886180578b92b969be7d9d756c84a1"
115
+ },
116
+ {
117
+ "bytes": 4223,
118
+ "path": "experiments/E002/conditions/c-residual-h16-f186e90df2/run_manifest.yaml",
119
+ "sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73"
120
+ },
121
+ {
122
+ "bytes": 9087,
123
+ "path": "experiments/E002/experiment-registry.yaml",
124
+ "sha256": "f9d72a714cac875ea7d73888f588cd5def5c12a699432895f312fd0c17b61bf4"
125
+ },
126
+ {
127
+ "bytes": 33817,
128
+ "path": "experiments/E002/experiment-summary.json",
129
+ "sha256": "f773c22dbbb72885897f8217fca205868e80274a4bafb6737698e83896f433ea"
130
+ },
131
+ {
132
+ "bytes": 2972,
133
+ "path": "experiments/E002/fresh-test-firewall-receipt.json",
134
+ "sha256": "54a05d0efacd2f3355f3d14d904c275635715b440c9d22ca49a0b88406c48084"
135
+ },
136
+ {
137
+ "bytes": 9747,
138
+ "path": "experiments/E002/fresh-test-summary.json",
139
+ "sha256": "3751e6ac03749b0bee791baedd4317a269747e35eb92e7b27cd38f2d76373136"
140
+ },
141
+ {
142
+ "bytes": 743,
143
+ "path": "experiments/E002/protocol-freeze-receipt.json",
144
+ "sha256": "982480bfa61e1c557302219ff1b8a7d379ed0161f72d74fea43171bd4b3c4518"
145
+ },
146
+ {
147
+ "bytes": 3366,
148
+ "path": "experiments/E002/runs/E002-R001/condition_config.yaml",
149
+ "sha256": "5adc301234cd1418e6c93b2628b2b7f6b869335b76ce5e20dd76e437cad74878"
150
+ },
151
+ {
152
+ "bytes": 10903,
153
+ "path": "experiments/E002/runs/E002-R001/evaluation_validation.json",
154
+ "sha256": "0fda3f48a9f3ac61cdc91d4e13c0cd0d7bbb006c23a4f074d0426313da506b06"
155
+ },
156
+ {
157
+ "bytes": 3518,
158
+ "path": "experiments/E002/runs/E002-R001/run_manifest.yaml",
159
+ "sha256": "d273fe41b8dfcbf020c4048aeec3e47296de4bf13bc56bef369df56a8d6c328b"
160
+ },
161
+ {
162
+ "bytes": 3539,
163
+ "path": "experiments/E002/runs/E002-R002/condition_config.yaml",
164
+ "sha256": "a01fabd1cc84b872251ab1edfbaa3ebbd3763559430046df5c0b07d63d9d441c"
165
+ },
166
+ {
167
+ "bytes": 12007,
168
+ "path": "experiments/E002/runs/E002-R002/evaluation_validation.json",
169
+ "sha256": "17a116944f1967d60d2427975690f635f4886180578b92b969be7d9d756c84a1"
170
+ },
171
+ {
172
+ "bytes": 4223,
173
+ "path": "experiments/E002/runs/E002-R002/run_manifest.yaml",
174
+ "sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73"
175
+ },
176
+ {
177
+ "bytes": 3575,
178
+ "path": "experiments/E002/runs/E002-R003/condition_config.yaml",
179
+ "sha256": "806740a24d00c5938c83458e923f5a919b8328052c816cc9f19cbaed5c628877"
180
+ },
181
+ {
182
+ "bytes": 12008,
183
+ "path": "experiments/E002/runs/E002-R003/evaluation_validation.json",
184
+ "sha256": "237728c0a5594821ef420a5ec4452943767845a7058ccf32f7dc6515ab7cd330"
185
+ },
186
+ {
187
+ "bytes": 4245,
188
+ "path": "experiments/E002/runs/E002-R003/run_manifest.yaml",
189
+ "sha256": "6ec30c7b61750729b9750a0fbb86e983c1192defd433e9986c69b6be3fb77298"
190
+ },
191
+ {
192
+ "bytes": 3573,
193
+ "path": "experiments/E002/runs/E002-R004/condition_config.yaml",
194
+ "sha256": "8b3b1638bf3f4527a3f65a786480482bdececdd4a3c2b5c6e37689fd42fd5bc6"
195
+ },
196
+ {
197
+ "bytes": 12025,
198
+ "path": "experiments/E002/runs/E002-R004/evaluation_validation.json",
199
+ "sha256": "7c451bf90cb7e2f1ea50a3dff1647bca8a9cea6d4adf9003733d8140e6beb741"
200
+ },
201
+ {
202
+ "bytes": 4247,
203
+ "path": "experiments/E002/runs/E002-R004/run_manifest.yaml",
204
+ "sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac"
205
+ },
206
+ {
207
+ "bytes": 4361,
208
+ "path": "experiments/E002/selection-validation.json",
209
+ "sha256": "8c3e64ef2f874fda8ec4d263d389dacb5abb1f5b2253bd4d47b9cce9fd7a7ec4"
210
  }
211
+ ],
212
+ "schema_version": "upvla.hf_package.v1",
213
+ "source_git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a"
214
  }
experiments/E002/E002-CARD.md ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # E002 — SE(3) B-spline / minimum-jerk recovery head
2
+
3
+ Status: **completed / rejected**
4
+ Selected on validation: **E002-R004 quintic minimum-jerk CP6**
5
+ Real Pi0.5 rollout authorized: **no**
6
+
7
+ ## Decision
8
+
9
+ The selected minimum-jerk head passed every validation gate. On the one-time fresh test it reduced group-macro command jerk from 3.1432 to 2.5615 (18.50%; paired ratio 0.815, 95% group-bootstrap CI [0.762, 0.869]) but missed the frozen 25% effect-size gate. Path score changed from 0.801884 to 0.835376 (4.18% worse), also missing the frozen 2% non-inferiority bound. Endpoint and latency gates passed. The hypothesis is therefore **rejected**, and E002 does not authorize a real Pi0.5 rollout.
10
+
11
+ ## Frozen mathematics
12
+
13
+ For normalized time `u=t/0.8` and the 16-point grid `t_k=0.05k s`, both structured heads predict local initial-tool-frame tangent controls `xi=[v_m, omega_rad]`. The residual transform and command use
14
+
15
+ `Delta T(u) = Exp_SE(3)(xi_res(u))`, `T_cmd[k] = T_base[k] @ Delta T(u_k)`.
16
+
17
+ - Cubic: six controls `[0,c1,...,c5]`, open-uniform clamped degree-3 knots `[0,0,0,0,1/3,2/3,1,1,1,1]`. The tangent curve is C2 at its simple interior knots and begins at the identity residual.
18
+ - Minimum jerk: degree-5 Bernstein controls `[0,0,0,q,q,q]`, exactly `xi_res(u)=(10u^3-15u^4+6u^5)q`. It has zero endpoint velocity and acceleration and is the integrated-squared-jerk minimizer under those boundary conditions.
19
+ - Translation controls are metres in the translational component of `se(3)`; rotation controls are radians in `so(3)`. Both are norm-bounded before reconstruction. Duration and confidence heads were excluded as unidentifiable/unsupervised.
20
+
21
+ Unit tests cover SE(3) exp/log, body-frame right composition, control-point identity initialization, B-spline partition/endpoints, and exact quintic minimum-jerk phase.
22
+
23
+ ## Validation (1,298 samples / 9 independent groups)
24
+
25
+ | Run | Condition | Path score | Endpoint mm | Endpoint deg | Command jerk | GPU p95 ms |
26
+ |---|---|---:|---:|---:|---:|---:|
27
+ | R001 | endpoint + profiled path | 0.873585 | 6.187 | 3.882 | 2.221 | 3.137 |
28
+ | R002 | residual H16 | 0.856968 | 6.062 | 3.858 | 3.137 | 8.201 |
29
+ | R003 | clamped cubic CP6 | 0.847769 | 6.003 | 3.811 | 2.313 | 8.216 |
30
+ | R004 | quintic minimum-jerk CP6 | 0.855793 | 6.209 | 3.745 | 2.272 | 8.980 |
31
+
32
+ R004 was selected by the frozen minimum-jerk-among-guarded-candidates rule. Its residual-H16 jerk ratio was 0.724, 95% CI [0.595, 0.808].
33
+
34
+ ## One-time fresh test (1,523 samples / 9 independent groups)
35
+
36
+ | Condition | Path score | Endpoint mm | Endpoint deg | Velocity RMSE | Accel RMSE | Jerk RMSE | Command jerk | GPU p95 ms |
37
+ |---|---:|---:|---:|---:|---:|---:|---:|---:|
38
+ | endpoint | 0.862117 | 6.095 | 3.800 | 0.03619 | 0.12567 | 3.063 | 2.537 | 3.137 |
39
+ | residual H16 | 0.801884 | 5.905 | 3.462 | 0.03316 | 0.14658 | 3.569 | 3.143 | 8.237 |
40
+ | minimum-jerk selected | 0.835376 | 5.905 | 3.701 | 0.03606 | 0.13714 | 3.082 | 2.562 | 8.214 |
41
+
42
+ Fresh-test failures: jerk reduction was 18.50% (<25%) and selected path was 4.18% worse than residual H16 (>2%). The paired jerk CI still excluded no improvement, so the architecture reduced jerk consistently but not by the pre-registered magnitude and not without excessive path tradeoff.
43
+
44
+ ## Data firewall and leakage
45
+
46
+ E002 uses exactly the 89 E001-train source groups, re-split to 71/9/9 groups (8,807/1,298/1,523 samples). All 22 old E001 validation/test groups are excluded. Both E002 holdouts cover RoboTwin and RLBench. The endpoint model was retrained from a fresh initialization on E002 train; proposal generation used `include_targets=False`, current observations/pose/task category only. Fresh test was opened once jointly for R001, R002, and validation-selected R004; R003 and the old E001 test were not opened as E002 test evidence.
47
+
48
+ ## Provenance
49
+
50
+ - Training code: `033cb092edc8a14430f35e5c78e06323ee14655a`
51
+ - Dataset revision: `77f5f306303361d602ab10759b3a139c67f2cf16`
52
+ - Split audit SHA-256: `022dac36e48ce65ea920a789a36e1458b40e78df7f55eabe15504fd8ef0f96be`
53
+ - Proposal manifest SHA-256: `7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3`
54
+ - Selection receipt SHA-256: `8c3e64ef2f874fda8ec4d263d389dacb5abb1f5b2253bd4d47b9cce9fd7a7ec4`
55
+ - Sealed firewall SHA-256: `54a05d0efacd2f3355f3d14d904c275635715b440c9d22ca49a0b88406c48084`
56
+ - Total training compute: 0.3595 GPU·h on NVIDIA RTX PRO 6000 Blackwell, Python 3.11.15, PyTorch 2.13.0+cu130, CUDA 13.0.
57
+
58
+ Machine-readable details, commands, environments, grouped CIs, hashes, and limitations are in `experiment-summary.json`.
experiments/E002/audits/action-target-audit.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "configured_limits": {
3
+ "step_rotation_rad": 0.2,
4
+ "step_translation_m": 0.01
5
+ },
6
+ "expert_success_counts": {
7
+ "recovery_task_success": 15000,
8
+ "source_expert_success": 15000
9
+ },
10
+ "git_commit": "91b8c9473cfc60c8bfce3848c900aba8e4594ca1",
11
+ "horizon_reaches_target_count": 0,
12
+ "horizon_seconds": 0.8,
13
+ "horizon_steps": 16,
14
+ "limit_violations": {
15
+ "step_rotation": 0,
16
+ "step_translation": 0
17
+ },
18
+ "sample_count": 15000,
19
+ "schema_version": "precision.action_chunk.target_audit.v1",
20
+ "source_families": {
21
+ "rlbench": 5220,
22
+ "robotwin_generic": 9780
23
+ },
24
+ "step_rotation_rad": {
25
+ "max": 0.010471975511966547,
26
+ "mean": 0.0043272625200924675,
27
+ "min": 0.000270797193020592,
28
+ "p50": 0.0037907259009449306,
29
+ "p95": 0.00872664625997162,
30
+ "p99": 0.010053904361015777
31
+ },
32
+ "step_translation_twist_m": {
33
+ "max": 0.0006000026630050273,
34
+ "mean": 0.0003779863550717578,
35
+ "min": 1.719141227820467e-05,
36
+ "p50": 0.0004000002360415781,
37
+ "p95": 0.000600000146014978,
38
+ "p99": 0.0006000012005080267
39
+ },
40
+ "trajectory_duration_s": {
41
+ "max": 20.731572870249437,
42
+ "mean": 6.3629103453163545,
43
+ "min": 1.2569543470285411,
44
+ "p50": 5.903310058639175,
45
+ "p95": 11.372092291086012,
46
+ "p99": 14.754663633027155
47
+ },
48
+ "waypoint_rotation_rad": {
49
+ "max": 0.16755160819145598,
50
+ "mean": 0.037987402154382004,
51
+ "min": 0.0004985226170931637,
52
+ "p50": 0.030744580054978578,
53
+ "p95": 0.09772211052589472,
54
+ "p99": 0.1255383529084257
55
+ },
56
+ "waypoint_translation_m": {
57
+ "max": 0.00960000000000014,
58
+ "mean": 0.0033021575027860745,
59
+ "min": 1.719141112148997e-05,
60
+ "p50": 0.0029999999999999884,
61
+ "p95": 0.007499999999999985,
62
+ "p99": 0.008399999999999987
63
+ }
64
+ }
experiments/E002/audits/data-split-audit.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checks": {
3
+ "both_holdouts_cover_rlbench_and_robotwin": true,
4
+ "eligible_pool_equals_e001_train": true,
5
+ "group_disjoint": true,
6
+ "old_e001_holdouts_excluded": true,
7
+ "sample_partition_complete": true
8
+ },
9
+ "condition_config": {
10
+ "path": "/home/coder/share/UP-VLA-se3-bspline-E002/configs/experiments/E002/R001-endpoint.yaml",
11
+ "sha256": "5adc301234cd1418e6c93b2628b2b7f6b869335b76ce5e20dd76e437cad74878"
12
+ },
13
+ "data": {
14
+ "extraction_receipt": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
15
+ "extraction_receipt_sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd",
16
+ "hf_repo_id": "Shiki42/up-vla-precision-recovery-15k",
17
+ "revision": "77f5f306303361d602ab10759b3a139c67f2cf16",
18
+ "variant_id": "validated-physical-recovery-15k-v1"
19
+ },
20
+ "eligibility": "exact E001 train groups only",
21
+ "eligible_groups": 89,
22
+ "eligible_samples": 11628,
23
+ "excluded_e001_validation_test_group_count": 22,
24
+ "excluded_e001_validation_test_groups": [
25
+ "rlbench::rlbench/charger_plug::datasets/rlbench_expert_validation/samples/plug_charger_in_power_supply/episode_000000/trajectory.npz",
26
+ "rlbench::rlbench/square_ring::datasets/rlbench_expert_validation/samples/insert_onto_square_peg/episode_000000/trajectory.npz",
27
+ "robotwin_generic::002_bowl/base3::candidate_0114",
28
+ "robotwin_generic::005_french-fries/base0::candidate_0002",
29
+ "robotwin_generic::006_hamburg/base2::candidate_0116",
30
+ "robotwin_generic::023_tissue-box/base0::candidate_0007",
31
+ "robotwin_generic::039_mug/base0::candidate_0010",
32
+ "robotwin_generic::039_mug/base10::candidate_0120",
33
+ "robotwin_generic::045_sand-clock/base3::candidate_0068",
34
+ "robotwin_generic::048_stapler/base1::candidate_0071",
35
+ "robotwin_generic::057_toycar/base1::candidate_0078",
36
+ "robotwin_generic::063_tabletrashbin/base3::candidate_0132",
37
+ "robotwin_generic::070_paymentsign/base0::candidate_0028",
38
+ "robotwin_generic::075_bread/base1::candidate_0086",
39
+ "robotwin_generic::080_pillbottle/base4::candidate_0090",
40
+ "robotwin_generic::081_playingcards/base1::candidate_0091",
41
+ "robotwin_generic::082_smallshovel/base0::candidate_0037",
42
+ "robotwin_generic::095_glue/base1::candidate_0099",
43
+ "robotwin_generic::099_fan/base3::candidate_0100",
44
+ "robotwin_generic::105_sauce-can/base0::candidate_0050",
45
+ "robotwin_generic::110_basket/base1::candidate_0108",
46
+ "robotwin_generic::112_tea-box/base3::candidate_0110"
47
+ ],
48
+ "experiment_id": "E002",
49
+ "hash_namespace": "E002-source-group-split-v1::",
50
+ "protocol": "e001-train-only-source-family-stratified-group-hash-e002-v1",
51
+ "schema_version": "precision.e002.split_audit.v1",
52
+ "splits": {
53
+ "test": {
54
+ "group_ids_sha256": "d07ef84deb50a218006c6bfe8d40e9819b29b08d5308b4ee0ecbea2ac31617d4",
55
+ "groups": 9,
56
+ "groups_by_source_family": {
57
+ "rlbench": 1,
58
+ "robotwin_generic": 8
59
+ },
60
+ "groups_sorted": [
61
+ "rlbench::rlbench/knife::datasets/rlbench_expert_validation/samples/put_knife_in_knife_block/episode_000000/trajectory.npz",
62
+ "robotwin_generic::023_tissue-box/base1::candidate_0065",
63
+ "robotwin_generic::049_shampoo/base1::candidate_0017",
64
+ "robotwin_generic::053_teanet/base5::candidate_0076",
65
+ "robotwin_generic::055_small-speaker/base1::candidate_0022",
66
+ "robotwin_generic::073_rubikscube/base0::candidate_0030",
67
+ "robotwin_generic::077_phone/base1::candidate_0087",
68
+ "robotwin_generic::107_soap/base0::candidate_0052",
69
+ "robotwin_generic::109_hydrating-oil/base1::candidate_0107"
70
+ ],
71
+ "sample_ids_sha256": "29c26a79a9031dc936a903297c80c2921e161a05045989d1f3f98a3513235b6f",
72
+ "samples": 1523,
73
+ "samples_by_source_family": {
74
+ "rlbench": 520,
75
+ "robotwin_generic": 1003
76
+ }
77
+ },
78
+ "train": {
79
+ "group_ids_sha256": "6a40fd0f8614d2e2210aeec603be405e1572405d292e12fc65fcbeba46c60572",
80
+ "groups": 71,
81
+ "groups_by_source_family": {
82
+ "rlbench": 7,
83
+ "robotwin_generic": 64
84
+ },
85
+ "groups_sorted": [
86
+ "rlbench::rlbench/cup::datasets/rlbench_expert_validation/samples/stack_cups/episode_000000/trajectory.npz",
87
+ "rlbench::rlbench/light_bulb::datasets/rlbench_expert_validation/samples/light_bulb_in/episode_000000/trajectory.npz",
88
+ "rlbench::rlbench/phone::datasets/rlbench_expert_validation/samples/phone_on_base/episode_000000/trajectory.npz",
89
+ "rlbench::rlbench/plate::datasets/rlbench_expert_validation/samples/take_plate_off_colored_dish_rack/episode_000000/trajectory.npz",
90
+ "rlbench::rlbench/screw_driver::datasets/rlbench_expert_validation/samples/screw_nail/episode_000000/trajectory.npz",
91
+ "rlbench::rlbench/toilet_roll::datasets/rlbench_expert_validation/samples/put_toilet_roll_on_stand/episode_000000/trajectory.npz",
92
+ "rlbench::rlbench/umbrella::datasets/rlbench_expert_validation/samples/put_umbrella_in_umbrella_stand/episode_000000/trajectory.npz",
93
+ "robotwin_generic::001_bottle/base0::candidate_0000",
94
+ "robotwin_generic::001_bottle/base10::candidate_0059",
95
+ "robotwin_generic::002_bowl/base2::candidate_0060",
96
+ "robotwin_generic::005_french-fries/base1::candidate_0061",
97
+ "robotwin_generic::006_hamburg/base0::candidate_0003",
98
+ "robotwin_generic::006_hamburg/base1::candidate_0062",
99
+ "robotwin_generic::018_microphone/base0::candidate_0005",
100
+ "robotwin_generic::021_cup/base0::candidate_0006",
101
+ "robotwin_generic::021_cup/base1::candidate_0064",
102
+ "robotwin_generic::023_tissue-box/base2::candidate_0119",
103
+ "robotwin_generic::024_scanner/base2::candidate_0008",
104
+ "robotwin_generic::039_mug/base1::candidate_0066",
105
+ "robotwin_generic::041_shoe/base0::candidate_0011",
106
+ "robotwin_generic::041_shoe/base2::candidate_0067",
107
+ "robotwin_generic::041_shoe/base4::candidate_0121",
108
+ "robotwin_generic::045_sand-clock/base0::candidate_0013",
109
+ "robotwin_generic::046_alarm-clock/base1::candidate_0014",
110
+ "robotwin_generic::046_alarm-clock/base2::candidate_0069",
111
+ "robotwin_generic::046_alarm-clock/base3::candidate_0122",
112
+ "robotwin_generic::047_mouse/base0::candidate_0015",
113
+ "robotwin_generic::047_mouse/base2::candidate_0123",
114
+ "robotwin_generic::048_stapler/base2::candidate_0124",
115
+ "robotwin_generic::049_shampoo/base3::candidate_0125",
116
+ "robotwin_generic::050_bell/base1::candidate_0073",
117
+ "robotwin_generic::051_candlestick/base1::candidate_0074",
118
+ "robotwin_generic::051_candlestick/base2::candidate_0126",
119
+ "robotwin_generic::052_dumbbell/base2::candidate_0075",
120
+ "robotwin_generic::052_dumbbell/base4::candidate_0127",
121
+ "robotwin_generic::053_teanet/base1::candidate_0021",
122
+ "robotwin_generic::053_teanet/base7::candidate_0128",
123
+ "robotwin_generic::055_small-speaker/base2::candidate_0077",
124
+ "robotwin_generic::057_toycar/base0::candidate_0023",
125
+ "robotwin_generic::057_toycar/base2::candidate_0129",
126
+ "robotwin_generic::059_pencup/base0::candidate_0024",
127
+ "robotwin_generic::062_plasticbox/base0::candidate_0025",
128
+ "robotwin_generic::062_plasticbox/base10::candidate_0131",
129
+ "robotwin_generic::062_plasticbox/base1::candidate_0080",
130
+ "robotwin_generic::063_tabletrashbin/base0::candidate_0026",
131
+ "robotwin_generic::068_boxdrink/base3::candidate_0082",
132
+ "robotwin_generic::070_paymentsign/base1::candidate_0083",
133
+ "robotwin_generic::071_can/base0::candidate_0029",
134
+ "robotwin_generic::073_rubikscube/base1::candidate_0085",
135
+ "robotwin_generic::077_phone/base0::candidate_0032",
136
+ "robotwin_generic::079_remotecontrol/base0::candidate_0034",
137
+ "robotwin_generic::079_remotecontrol/base1::candidate_0089",
138
+ "robotwin_generic::080_pillbottle/base1::candidate_0035",
139
+ "robotwin_generic::082_smallshovel/base1::candidate_0092",
140
+ "robotwin_generic::084_woodenmallet/base3::candidate_0038",
141
+ "robotwin_generic::091_kettle/base0::candidate_0042",
142
+ "robotwin_generic::093_brush-pen/base1::candidate_0098",
143
+ "robotwin_generic::099_fan/base2::candidate_0046",
144
+ "robotwin_generic::100_seal/base1::candidate_0047",
145
+ "robotwin_generic::100_seal/base2::candidate_0101",
146
+ "robotwin_generic::101_milk-tea/base1::candidate_0102",
147
+ "robotwin_generic::102_roller/base1::candidate_0049",
148
+ "robotwin_generic::105_sauce-can/base2::candidate_0104",
149
+ "robotwin_generic::106_skillet/base0::candidate_0051",
150
+ "robotwin_generic::106_skillet/base1::candidate_0105",
151
+ "robotwin_generic::110_basket/base0::candidate_0054",
152
+ "robotwin_generic::111_callbell/base1::candidate_0055",
153
+ "robotwin_generic::113_coffee-box/base0::candidate_0057",
154
+ "robotwin_generic::113_coffee-box/base1::candidate_0111",
155
+ "robotwin_generic::114_bottle/base1::candidate_0058",
156
+ "robotwin_generic::114_bottle/base2::candidate_0112"
157
+ ],
158
+ "sample_ids_sha256": "3ac6deb402131311ce3284e053a3e7ffcf8e5ff8485eaf1ca3359fc14122b2df",
159
+ "samples": 8807,
160
+ "samples_by_source_family": {
161
+ "rlbench": 3140,
162
+ "robotwin_generic": 5667
163
+ }
164
+ },
165
+ "validation": {
166
+ "group_ids_sha256": "60512895da2b7640139d04845c3d8fc0e67bd0f1a628215789ba11ab5a6071b2",
167
+ "groups": 9,
168
+ "groups_by_source_family": {
169
+ "rlbench": 1,
170
+ "robotwin_generic": 8
171
+ },
172
+ "groups_sorted": [
173
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz",
174
+ "robotwin_generic::002_bowl/base1::candidate_0001",
175
+ "robotwin_generic::050_bell/base0::candidate_0018",
176
+ "robotwin_generic::051_candlestick/base0::candidate_0019",
177
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081",
178
+ "robotwin_generic::068_boxdrink/base2::candidate_0027",
179
+ "robotwin_generic::071_can/base1::candidate_0084",
180
+ "robotwin_generic::107_soap/base1::candidate_0106",
181
+ "robotwin_generic::111_callbell/base2::candidate_0109"
182
+ ],
183
+ "sample_ids_sha256": "bb16c31fbf96333a43f2aa8ecee6aa774a51c9f646a5b22c3667b534f805ba72",
184
+ "samples": 1298,
185
+ "samples_by_source_family": {
186
+ "rlbench": 520,
187
+ "robotwin_generic": 778
188
+ }
189
+ }
190
+ }
191
+ }
experiments/E002/audits/proposal-manifest.json ADDED
The diff for this file is too large to render. See raw diff
 
experiments/E002/conditions/c-endpoint-5d2dff166d/CONDITION_CARD.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Endpoint baseline + profiled controller
2
+
3
+ - Experiment / run: E002 / E002-R001
4
+ - Condition: c-endpoint-5d2dff166d
5
+ - Role: baseline
6
+ - Checkpoint SHA-256: 0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da
7
+ - Code commit: 033cb092edc8a14430f35e5c78e06323ee14655a (clean)
8
+ - Dataset: Shiki42/up-vla-precision-recovery-15k @ 77f5f306303361d602ab10759b3a139c67f2cf16
9
+ - E002 split: E001-train-only grouped re-split; train 8,807/71 groups, validation 1,298/9, fresh test 1,523/9
10
+ - Geometry: local tool frame, right SE(3) composition. Endpoint prediction converted to a deterministic profiled 16-waypoint path.
11
+ - Validation: path 0.873585, endpoint 6.187 mm / 3.882 deg, command jerk 2.221
12
+ - Fresh test: path 0.862117, endpoint 6.095 mm / 3.800 deg, command jerk 2.537, GPU p95 3.137 ms
13
+ - Decision: E002 rejected; real Pi0.5 rollout not authorized
14
+
15
+ See sibling resolved config, immutable run manifest, validation report, and one-time fresh-test report.
experiments/E002/conditions/c-endpoint-5d2dff166d/best.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da
3
+ size 189704503
experiments/E002/conditions/c-endpoint-5d2dff166d/config_resolved.yaml ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
3
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
4
+ variant_id: validated-physical-recovery-15k-v1
5
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
6
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
7
+ sample_count: 15000
8
+ eligible_sample_count: 11628
9
+ eligible_group_count: 89
10
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
11
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
12
+ split_namespace: 'E002-source-group-split-v1::'
13
+ validation_fraction: 0.1
14
+ test_fraction: 0.1
15
+ geometry:
16
+ cumulative_waypoint_representation: current_tool_relative_pose
17
+ step_representation: integrated_body_twist
18
+ step_integration: right_se3_exponential
19
+ residual_curve_frame: initial_local_tool_frame
20
+ residual_composition: base_right_multiply_se3_exponential
21
+ structured_translation: se3_tangent_translation_control_points_m
22
+ structured_rotation: so3_lie_algebra_control_points_rad
23
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
24
+ residual_boundary_at_t0: identity
25
+ model:
26
+ backbone:
27
+ hidden_size: 256
28
+ fusion_layers: 4
29
+ fusion_heads: 8
30
+ fusion_mlp_ratio: 4
31
+ dropout: 0.1
32
+ vision_backend: conv
33
+ text_backend: embedding
34
+ pretrained_model_name: google/siglip-base-patch16-224
35
+ freeze_pretrained: true
36
+ image_size: 224
37
+ vision_tokens: 64
38
+ depth_tokens: 36
39
+ point_tokens: 32
40
+ point_samples: 256
41
+ max_text_tokens: 32
42
+ text_vocab_size: 32768
43
+ max_joints: 7
44
+ aloha_joints: 6
45
+ panda_joints: 7
46
+ translation_scale_m: 0.05
47
+ rotation_scale_rad: 0.5235987755982988
48
+ joint_scale_rad: 0.5
49
+ horizon_steps: 16
50
+ horizon_seconds: 0.8
51
+ decoder_layers: 4
52
+ decoder_heads: 8
53
+ decoder_mlp_ratio: 4
54
+ control_points: 6
55
+ maximum_step_translation_m: 0.01
56
+ maximum_step_rotation_rad: 0.2
57
+ maximum_residual_translation_m: 0.05
58
+ maximum_residual_rotation_rad: 0.52
59
+ action_head: endpoint
60
+ base_conditioning: none
61
+ trajectory_basis: none
62
+ parameter_counts:
63
+ total: 15790420
64
+ trainable: 15790420
65
+ frozen: 0
66
+ loss:
67
+ translation: 1.0
68
+ rotation_vector: 1.0
69
+ rotation_geodesic: 0.25
70
+ joint: 0.5
71
+ gripper: 0.1
72
+ training:
73
+ seed: 42
74
+ batch_size: 64
75
+ gradient_accumulation_steps: 1
76
+ epochs: 100
77
+ max_steps: 3000
78
+ learning_rate: 0.0003
79
+ weight_decay: 0.05
80
+ warmup_fraction: 0.05
81
+ gradient_clip_norm: 1.0
82
+ precision: bf16
83
+ workers: 8
84
+ log_every_steps: 10
85
+ validation_every_steps: 250
86
+ checkpoint_every_steps: 500
87
+ evaluation:
88
+ version: se3-spline-eval-v1
89
+ trajectory_loss:
90
+ path_translation: 1.0
91
+ path_rotation: 1.0
92
+ endpoint_translation: 2.0
93
+ endpoint_rotation: 2.0
94
+ velocity: 1.0
95
+ acceleration: 0.1
96
+ jerk: 0.05
97
+ residual_magnitude: 0.02
98
+ selection_split: validation
99
+ test_firewall_opened: false
100
+ group_bootstrap_seed: 20260810
101
+ group_bootstrap_resamples: 10000
102
+ runtime:
103
+ data_root: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples
104
+ base_proposal_root: null
105
+ output_directory: /home/coder/share/experiment-runs/E002/E002-R001
106
+ split_counts:
107
+ train_samples: 8807
108
+ validation_samples: 1298
109
+ train_groups: 71
110
+ validation_groups: 9
111
+ overlap: false
experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_fresh_test.json ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": null,
7
+ "checkpoint_run": null,
8
+ "condition_id": "c-endpoint-5d2dff166d",
9
+ "data_extraction_receipt": {
10
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
11
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
12
+ },
13
+ "experiment_id": "E002",
14
+ "group_bootstrap_95ci": {
15
+ "confidence": 0.95,
16
+ "group_count": 9,
17
+ "intervals": {
18
+ "acceleration_rmse": {
19
+ "lower": 0.1160274123152097,
20
+ "upper": 0.13443885842959086
21
+ },
22
+ "command_acceleration_rms": {
23
+ "lower": 0.09190107720593611,
24
+ "upper": 0.1158909913773338
25
+ },
26
+ "command_jerk_rms": {
27
+ "lower": 2.229633437262641,
28
+ "upper": 2.815442830324173
29
+ },
30
+ "endpoint_rotation_rmse_deg": {
31
+ "lower": 3.132349097728729,
32
+ "upper": 4.655834012561375
33
+ },
34
+ "endpoint_translation_rmse_mm": {
35
+ "lower": 5.6727268086539375,
36
+ "upper": 6.541422454516093
37
+ },
38
+ "group_macro_path_score": {
39
+ "lower": 0.7736077094740339,
40
+ "upper": 0.9841096719768314
41
+ },
42
+ "jerk_rmse": {
43
+ "lower": 2.810545167658064,
44
+ "upper": 3.28650067448616
45
+ },
46
+ "path_rotation_rmse_deg": {
47
+ "lower": 2.019929062657886,
48
+ "upper": 2.9412611941496527
49
+ },
50
+ "path_translation_rmse_mm": {
51
+ "lower": 3.5075451268090143,
52
+ "upper": 4.077366487847434
53
+ },
54
+ "residual_rotation_mean_deg": {
55
+ "lower": 0.0,
56
+ "upper": 0.0
57
+ },
58
+ "residual_translation_mean_mm": {
59
+ "lower": 0.0,
60
+ "upper": 0.0
61
+ },
62
+ "velocity_rmse": {
63
+ "lower": 0.03070427512543069,
64
+ "upper": 0.043454261496663094
65
+ }
66
+ },
67
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
68
+ "resamples": 10000,
69
+ "seed": 20260810
70
+ },
71
+ "group_count": 9,
72
+ "grouped_metrics": {
73
+ "acceleration_rmse": 0.12567059695720673,
74
+ "command_acceleration_rms": 0.10415160655975342,
75
+ "command_jerk_rms": 2.5366859436035156,
76
+ "endpoint_rotation_rmse_deg": 3.7998945713043213,
77
+ "endpoint_translation_rmse_mm": 6.095478534698486,
78
+ "jerk_rmse": 3.062636137008667,
79
+ "path_rotation_rmse_deg": 2.4173941612243652,
80
+ "path_translation_rmse_mm": 3.786381721496582,
81
+ "residual_rotation_mean_deg": 0.0,
82
+ "residual_translation_mean_mm": 0.0,
83
+ "velocity_rmse": 0.03618572652339935
84
+ },
85
+ "latency": null,
86
+ "losses": {
87
+ "acceleration": 0.0009877999546006322,
88
+ "endpoint_rotation": 0.1428782343864441,
89
+ "endpoint_translation": 0.002669057110324502,
90
+ "jerk": 0.0014431423041969538,
91
+ "path_rotation": 0.08060792833566666,
92
+ "path_translation": 0.001019860152155161,
93
+ "residual_magnitude": 0.0,
94
+ "total": 0.4344918727874756,
95
+ "velocity": 0.06159853935241699
96
+ },
97
+ "metrics": {
98
+ "acceleration_rmse": 0.13223324716091156,
99
+ "command_acceleration_rms": 0.113448865711689,
100
+ "command_jerk_rms": 2.734924793243408,
101
+ "endpoint_rotation_rmse_deg": 4.957154750823975,
102
+ "endpoint_translation_rmse_mm": 6.327389240264893,
103
+ "jerk_rmse": 3.2046091556549072,
104
+ "path_rotation_rmse_deg": 3.1238508224487305,
105
+ "path_translation_rmse_mm": 3.9112534523010254,
106
+ "residual_rotation_mean_deg": 0.0,
107
+ "residual_translation_mean_mm": 0.0,
108
+ "velocity_rmse": 0.04589399695396423
109
+ },
110
+ "per_group_metrics": {
111
+ "rlbench::rlbench/knife::datasets/rlbench_expert_validation/samples/put_knife_in_knife_block/episode_000000/trajectory.npz": {
112
+ "acceleration_rmse": 0.14491471648216248,
113
+ "command_acceleration_rms": 0.13358986377716064,
114
+ "command_jerk_rms": 3.1502559185028076,
115
+ "endpoint_rotation_rmse_deg": 6.857841491699219,
116
+ "endpoint_translation_rmse_mm": 7.155666828155518,
117
+ "jerk_rmse": 3.4326634407043457,
118
+ "path_rotation_rmse_deg": 4.303040027618408,
119
+ "path_translation_rmse_mm": 4.401074409484863,
120
+ "residual_rotation_mean_deg": 0.0,
121
+ "residual_translation_mean_mm": 0.0,
122
+ "velocity_rmse": 0.06246773898601532
123
+ },
124
+ "robotwin_generic::023_tissue-box/base1::candidate_0065": {
125
+ "acceleration_rmse": 0.12678000330924988,
126
+ "command_acceleration_rms": 0.11714279651641846,
127
+ "command_jerk_rms": 2.880620002746582,
128
+ "endpoint_rotation_rmse_deg": 2.1245810985565186,
129
+ "endpoint_translation_rmse_mm": 6.784346103668213,
130
+ "jerk_rmse": 3.0708529949188232,
131
+ "path_rotation_rmse_deg": 1.4709573984146118,
132
+ "path_translation_rmse_mm": 4.327749729156494,
133
+ "residual_rotation_mean_deg": 0.0,
134
+ "residual_translation_mean_mm": 0.0,
135
+ "velocity_rmse": 0.02322014793753624
136
+ },
137
+ "robotwin_generic::049_shampoo/base1::candidate_0017": {
138
+ "acceleration_rmse": 0.10978566110134125,
139
+ "command_acceleration_rms": 0.07980752736330032,
140
+ "command_jerk_rms": 2.0340023040771484,
141
+ "endpoint_rotation_rmse_deg": 3.5188851356506348,
142
+ "endpoint_translation_rmse_mm": 5.774036407470703,
143
+ "jerk_rmse": 2.7564311027526855,
144
+ "path_rotation_rmse_deg": 2.219775438308716,
145
+ "path_translation_rmse_mm": 3.5804269313812256,
146
+ "residual_rotation_mean_deg": 0.0,
147
+ "residual_translation_mean_mm": 0.0,
148
+ "velocity_rmse": 0.03317262604832649
149
+ },
150
+ "robotwin_generic::053_teanet/base5::candidate_0076": {
151
+ "acceleration_rmse": 0.14248709380626678,
152
+ "command_acceleration_rms": 0.11345230042934418,
153
+ "command_jerk_rms": 2.8445591926574707,
154
+ "endpoint_rotation_rmse_deg": 4.174556732177734,
155
+ "endpoint_translation_rmse_mm": 6.955822467803955,
156
+ "jerk_rmse": 3.597935199737549,
157
+ "path_rotation_rmse_deg": 2.668811798095703,
158
+ "path_translation_rmse_mm": 4.369506359100342,
159
+ "residual_rotation_mean_deg": 0.0,
160
+ "residual_translation_mean_mm": 0.0,
161
+ "velocity_rmse": 0.03951823338866234
162
+ },
163
+ "robotwin_generic::055_small-speaker/base1::candidate_0022": {
164
+ "acceleration_rmse": 0.11616356670856476,
165
+ "command_acceleration_rms": 0.08757305145263672,
166
+ "command_jerk_rms": 2.1193184852600098,
167
+ "endpoint_rotation_rmse_deg": 3.386752128601074,
168
+ "endpoint_translation_rmse_mm": 5.382193565368652,
169
+ "jerk_rmse": 2.8224096298217773,
170
+ "path_rotation_rmse_deg": 2.141021966934204,
171
+ "path_translation_rmse_mm": 3.31095290184021,
172
+ "residual_rotation_mean_deg": 0.0,
173
+ "residual_translation_mean_mm": 0.0,
174
+ "velocity_rmse": 0.03220295161008835
175
+ },
176
+ "robotwin_generic::073_rubikscube/base0::candidate_0030": {
177
+ "acceleration_rmse": 0.1296936422586441,
178
+ "command_acceleration_rms": 0.11092998087406158,
179
+ "command_jerk_rms": 2.719470262527466,
180
+ "endpoint_rotation_rmse_deg": 3.1834702491760254,
181
+ "endpoint_translation_rmse_mm": 5.193059921264648,
182
+ "jerk_rmse": 3.1652793884277344,
183
+ "path_rotation_rmse_deg": 2.0245392322540283,
184
+ "path_translation_rmse_mm": 3.2024216651916504,
185
+ "residual_rotation_mean_deg": 0.0,
186
+ "residual_translation_mean_mm": 0.0,
187
+ "velocity_rmse": 0.031097741797566414
188
+ },
189
+ "robotwin_generic::077_phone/base1::candidate_0087": {
190
+ "acceleration_rmse": 0.1338159739971161,
191
+ "command_acceleration_rms": 0.1163584515452385,
192
+ "command_jerk_rms": 2.833106517791748,
193
+ "endpoint_rotation_rmse_deg": 3.4720327854156494,
194
+ "endpoint_translation_rmse_mm": 5.668212890625,
195
+ "jerk_rmse": 3.2681198120117188,
196
+ "path_rotation_rmse_deg": 2.2253170013427734,
197
+ "path_translation_rmse_mm": 3.4989163875579834,
198
+ "residual_rotation_mean_deg": 0.0,
199
+ "residual_translation_mean_mm": 0.0,
200
+ "velocity_rmse": 0.03366128355264664
201
+ },
202
+ "robotwin_generic::107_soap/base0::candidate_0052": {
203
+ "acceleration_rmse": 0.129174143075943,
204
+ "command_acceleration_rms": 0.10575971007347107,
205
+ "command_jerk_rms": 2.573106050491333,
206
+ "endpoint_rotation_rmse_deg": 3.802366256713867,
207
+ "endpoint_translation_rmse_mm": 5.7494964599609375,
208
+ "jerk_rmse": 3.157802104949951,
209
+ "path_rotation_rmse_deg": 2.411280393600464,
210
+ "path_translation_rmse_mm": 3.5700595378875732,
211
+ "residual_rotation_mean_deg": 0.0,
212
+ "residual_translation_mean_mm": 0.0,
213
+ "velocity_rmse": 0.03605728968977928
214
+ },
215
+ "robotwin_generic::109_hydrating-oil/base1::candidate_0107": {
216
+ "acceleration_rmse": 0.09822054207324982,
217
+ "command_acceleration_rms": 0.07275073975324631,
218
+ "command_jerk_rms": 1.675735354423523,
219
+ "endpoint_rotation_rmse_deg": 3.678565263748169,
220
+ "endpoint_translation_rmse_mm": 6.196468830108643,
221
+ "jerk_rmse": 2.2922325134277344,
222
+ "path_rotation_rmse_deg": 2.291806221008301,
223
+ "path_translation_rmse_mm": 3.816323757171631,
224
+ "residual_rotation_mean_deg": 0.0,
225
+ "residual_translation_mean_mm": 0.0,
226
+ "velocity_rmse": 0.03427349030971527
227
+ }
228
+ },
229
+ "prediction_source": "base_proposal",
230
+ "primary_path_score": {
231
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
232
+ "group_macro": 0.8621170043945312,
233
+ "micro": 1.0158955097198485
234
+ },
235
+ "run_id": "E002-R001",
236
+ "sample_count": 1523,
237
+ "schema_version": "precision.action_chunk.evaluation.v1",
238
+ "source_family_metrics": {
239
+ "rlbench": {
240
+ "group_count": 1,
241
+ "metrics": {
242
+ "acceleration_rmse": 0.14491471648216248,
243
+ "command_acceleration_rms": 0.13358986377716064,
244
+ "command_jerk_rms": 3.1502559185028076,
245
+ "endpoint_rotation_rmse_deg": 6.857841491699219,
246
+ "endpoint_translation_rmse_mm": 7.155666828155518,
247
+ "jerk_rmse": 3.4326634407043457,
248
+ "path_rotation_rmse_deg": 4.303040027618408,
249
+ "path_translation_rmse_mm": 4.401074409484863,
250
+ "residual_rotation_mean_deg": 0.0,
251
+ "residual_translation_mean_mm": 0.0,
252
+ "velocity_rmse": 0.06246773898601532
253
+ },
254
+ "path_score": 1.300715446472168,
255
+ "sample_count": 520
256
+ },
257
+ "robotwin_generic": {
258
+ "group_count": 8,
259
+ "metrics": {
260
+ "acceleration_rmse": 0.12515383958816528,
261
+ "command_acceleration_rms": 0.10144490003585815,
262
+ "command_jerk_rms": 2.492504596710205,
263
+ "endpoint_rotation_rmse_deg": 3.5959532260894775,
264
+ "endpoint_translation_rmse_mm": 5.852010250091553,
265
+ "jerk_rmse": 3.079735517501831,
266
+ "path_rotation_rmse_deg": 2.284306287765503,
267
+ "path_translation_rmse_mm": 3.631394863128662,
268
+ "residual_rotation_mean_deg": 0.0,
269
+ "residual_translation_mean_mm": 0.0,
270
+ "velocity_rmse": 0.03428050875663757
271
+ },
272
+ "path_score": 0.8200007438659668,
273
+ "sample_count": 1003
274
+ }
275
+ },
276
+ "split": "test",
277
+ "test_firewall_opened": true
278
+ }
experiments/E002/conditions/c-endpoint-5d2dff166d/evaluation_validation.json ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": null,
7
+ "checkpoint_run": null,
8
+ "condition_id": "c-endpoint-5d2dff166d",
9
+ "data_extraction_receipt": {
10
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
11
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
12
+ },
13
+ "experiment_id": "E002",
14
+ "group_bootstrap_95ci": {
15
+ "confidence": 0.95,
16
+ "group_count": 9,
17
+ "intervals": {
18
+ "acceleration_rmse": {
19
+ "lower": 0.08032658506029595,
20
+ "upper": 0.1295631282652418
21
+ },
22
+ "command_acceleration_rms": {
23
+ "lower": 0.06541380361498643,
24
+ "upper": 0.10663357377052306
25
+ },
26
+ "command_jerk_rms": {
27
+ "lower": 1.6106185335252021,
28
+ "upper": 2.6380579706695344
29
+ },
30
+ "endpoint_rotation_rmse_deg": {
31
+ "lower": 3.3786309235625795,
32
+ "upper": 4.45805601477623
33
+ },
34
+ "endpoint_translation_rmse_mm": {
35
+ "lower": 5.774903562333849,
36
+ "upper": 6.642621228430006
37
+ },
38
+ "group_macro_path_score": {
39
+ "lower": 0.8061851580937703,
40
+ "upper": 0.9451770804325738
41
+ },
42
+ "jerk_rmse": {
43
+ "lower": 1.995983472549253,
44
+ "upper": 3.205968793233236
45
+ },
46
+ "path_rotation_rmse_deg": {
47
+ "lower": 2.1476370354493457,
48
+ "upper": 2.7698281317949296
49
+ },
50
+ "path_translation_rmse_mm": {
51
+ "lower": 3.59573735859659,
52
+ "upper": 4.142488827308019
53
+ },
54
+ "residual_rotation_mean_deg": {
55
+ "lower": 0.0,
56
+ "upper": 0.0
57
+ },
58
+ "residual_translation_mean_mm": {
59
+ "lower": 0.0,
60
+ "upper": 0.0
61
+ },
62
+ "velocity_rmse": {
63
+ "lower": 0.032407565476993715,
64
+ "upper": 0.04119525947090652
65
+ }
66
+ },
67
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
68
+ "resamples": 10000,
69
+ "seed": 20260810
70
+ },
71
+ "group_count": 9,
72
+ "grouped_metrics": {
73
+ "acceleration_rmse": 0.11023411154747009,
74
+ "command_acceleration_rms": 0.09015228599309921,
75
+ "command_jerk_rms": 2.220862865447998,
76
+ "endpoint_rotation_rmse_deg": 3.881985664367676,
77
+ "endpoint_translation_rmse_mm": 6.186707973480225,
78
+ "jerk_rmse": 2.729586601257324,
79
+ "path_rotation_rmse_deg": 2.441077709197998,
80
+ "path_translation_rmse_mm": 3.853693962097168,
81
+ "residual_rotation_mean_deg": 0.0,
82
+ "residual_translation_mean_mm": 0.0,
83
+ "velocity_rmse": 0.036509882658720016
84
+ },
85
+ "latency": null,
86
+ "losses": {
87
+ "acceleration": 0.001078446046449244,
88
+ "endpoint_rotation": 0.11553905159235,
89
+ "endpoint_translation": 0.002643197774887085,
90
+ "jerk": 0.0016622950788587332,
91
+ "path_rotation": 0.06584077328443527,
92
+ "path_translation": 0.0010324910981580615,
93
+ "residual_magnitude": 0.0,
94
+ "total": 0.357170432806015,
95
+ "velocity": 0.053741686046123505
96
+ },
97
+ "metrics": {
98
+ "acceleration_rmse": 0.12457380443811417,
99
+ "command_acceleration_rms": 0.10650406777858734,
100
+ "command_jerk_rms": 2.6437199115753174,
101
+ "endpoint_rotation_rmse_deg": 3.9696099758148193,
102
+ "endpoint_translation_rmse_mm": 6.296663284301758,
103
+ "jerk_rmse": 3.105485677719116,
104
+ "path_rotation_rmse_deg": 2.5149660110473633,
105
+ "path_translation_rmse_mm": 3.9353995323181152,
106
+ "residual_rotation_mean_deg": 0.0,
107
+ "residual_translation_mean_mm": 0.0,
108
+ "velocity_rmse": 0.03739766776561737
109
+ },
110
+ "per_group_metrics": {
111
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
112
+ "acceleration_rmse": 0.12432977557182312,
113
+ "command_acceleration_rms": 0.11089512705802917,
114
+ "command_jerk_rms": 2.782877206802368,
115
+ "endpoint_rotation_rmse_deg": 4.350729465484619,
116
+ "endpoint_translation_rmse_mm": 6.224776744842529,
117
+ "jerk_rmse": 3.11281418800354,
118
+ "path_rotation_rmse_deg": 2.7672295570373535,
119
+ "path_translation_rmse_mm": 3.89414119720459,
120
+ "residual_rotation_mean_deg": 0.0,
121
+ "residual_translation_mean_mm": 0.0,
122
+ "velocity_rmse": 0.040785618126392365
123
+ },
124
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
125
+ "acceleration_rmse": 0.14718665182590485,
126
+ "command_acceleration_rms": 0.12139102071523666,
127
+ "command_jerk_rms": 2.9899542331695557,
128
+ "endpoint_rotation_rmse_deg": 4.108950138092041,
129
+ "endpoint_translation_rmse_mm": 7.592531204223633,
130
+ "jerk_rmse": 3.647094488143921,
131
+ "path_rotation_rmse_deg": 2.6102445125579834,
132
+ "path_translation_rmse_mm": 4.781116962432861,
133
+ "residual_rotation_mean_deg": 0.0,
134
+ "residual_translation_mean_mm": 0.0,
135
+ "velocity_rmse": 0.03910675644874573
136
+ },
137
+ "robotwin_generic::050_bell/base0::candidate_0018": {
138
+ "acceleration_rmse": 0.11602437496185303,
139
+ "command_acceleration_rms": 0.09816119074821472,
140
+ "command_jerk_rms": 2.3722805976867676,
141
+ "endpoint_rotation_rmse_deg": 3.9214015007019043,
142
+ "endpoint_translation_rmse_mm": 5.605432510375977,
143
+ "jerk_rmse": 2.9190683364868164,
144
+ "path_rotation_rmse_deg": 2.4512875080108643,
145
+ "path_translation_rmse_mm": 3.457277774810791,
146
+ "residual_rotation_mean_deg": 0.0,
147
+ "residual_translation_mean_mm": 0.0,
148
+ "velocity_rmse": 0.03643663600087166
149
+ },
150
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
151
+ "acceleration_rmse": 0.0005567761836573482,
152
+ "command_acceleration_rms": 0.00045566412154585123,
153
+ "command_jerk_rms": 0.017189867794513702,
154
+ "endpoint_rotation_rmse_deg": 5.807064056396484,
155
+ "endpoint_translation_rmse_mm": 6.1670756340026855,
156
+ "jerk_rmse": 0.01718975231051445,
157
+ "path_rotation_rmse_deg": 3.5094783306121826,
158
+ "path_translation_rmse_mm": 3.727048635482788,
159
+ "residual_rotation_mean_deg": 0.0,
160
+ "residual_translation_mean_mm": 0.0,
161
+ "velocity_rmse": 0.05181599035859108
162
+ },
163
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
164
+ "acceleration_rmse": 0.12440907955169678,
165
+ "command_acceleration_rms": 0.09963987022638321,
166
+ "command_jerk_rms": 2.5165276527404785,
167
+ "endpoint_rotation_rmse_deg": 2.6556525230407715,
168
+ "endpoint_translation_rmse_mm": 6.43054723739624,
169
+ "jerk_rmse": 3.1067402362823486,
170
+ "path_rotation_rmse_deg": 1.726789116859436,
171
+ "path_translation_rmse_mm": 4.088305950164795,
172
+ "residual_rotation_mean_deg": 0.0,
173
+ "residual_translation_mean_mm": 0.0,
174
+ "velocity_rmse": 0.02654867060482502
175
+ },
176
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
177
+ "acceleration_rmse": 0.13240712881088257,
178
+ "command_acceleration_rms": 0.08462731540203094,
179
+ "command_jerk_rms": 1.9978476762771606,
180
+ "endpoint_rotation_rmse_deg": 3.0848066806793213,
181
+ "endpoint_translation_rmse_mm": 5.198719024658203,
182
+ "jerk_rmse": 3.147235155105591,
183
+ "path_rotation_rmse_deg": 1.9724689722061157,
184
+ "path_translation_rmse_mm": 3.3021345138549805,
185
+ "residual_rotation_mean_deg": 0.0,
186
+ "residual_translation_mean_mm": 0.0,
187
+ "velocity_rmse": 0.030027437955141068
188
+ },
189
+ "robotwin_generic::071_can/base1::candidate_0084": {
190
+ "acceleration_rmse": 0.11730268597602844,
191
+ "command_acceleration_rms": 0.10736808180809021,
192
+ "command_jerk_rms": 2.672065019607544,
193
+ "endpoint_rotation_rmse_deg": 3.5667669773101807,
194
+ "endpoint_translation_rmse_mm": 6.416263580322266,
195
+ "jerk_rmse": 2.9105913639068604,
196
+ "path_rotation_rmse_deg": 2.2308413982391357,
197
+ "path_translation_rmse_mm": 3.9885590076446533,
198
+ "residual_rotation_mean_deg": 0.0,
199
+ "residual_translation_mean_mm": 0.0,
200
+ "velocity_rmse": 0.03380422294139862
201
+ },
202
+ "robotwin_generic::107_soap/base1::candidate_0106": {
203
+ "acceleration_rmse": 0.11784116178750992,
204
+ "command_acceleration_rms": 0.09079129248857498,
205
+ "command_jerk_rms": 2.2749669551849365,
206
+ "endpoint_rotation_rmse_deg": 3.645209312438965,
207
+ "endpoint_translation_rmse_mm": 5.499370574951172,
208
+ "jerk_rmse": 2.957634449005127,
209
+ "path_rotation_rmse_deg": 2.308201313018799,
210
+ "path_translation_rmse_mm": 3.419480800628662,
211
+ "residual_rotation_mean_deg": 0.0,
212
+ "residual_translation_mean_mm": 0.0,
213
+ "velocity_rmse": 0.0345214307308197
214
+ },
215
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
216
+ "acceleration_rmse": 0.11204928904771805,
217
+ "command_acceleration_rms": 0.09804097563028336,
218
+ "command_jerk_rms": 2.3640592098236084,
219
+ "endpoint_rotation_rmse_deg": 3.7972915172576904,
220
+ "endpoint_translation_rmse_mm": 6.545653343200684,
221
+ "jerk_rmse": 2.747910737991333,
222
+ "path_rotation_rmse_deg": 2.393160104751587,
223
+ "path_translation_rmse_mm": 4.025181770324707,
224
+ "residual_rotation_mean_deg": 0.0,
225
+ "residual_translation_mean_mm": 0.0,
226
+ "velocity_rmse": 0.035542167723178864
227
+ }
228
+ },
229
+ "prediction_source": "base_proposal",
230
+ "primary_path_score": {
231
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
232
+ "group_macro": 0.8735849380493164,
233
+ "micro": 0.8965331554412841
234
+ },
235
+ "run_id": "E002-R001",
236
+ "sample_count": 1298,
237
+ "schema_version": "precision.action_chunk.evaluation.v1",
238
+ "source_family_metrics": {
239
+ "rlbench": {
240
+ "group_count": 1,
241
+ "metrics": {
242
+ "acceleration_rmse": 0.12432977557182312,
243
+ "command_acceleration_rms": 0.11089512705802917,
244
+ "command_jerk_rms": 2.782877206802368,
245
+ "endpoint_rotation_rmse_deg": 4.350729465484619,
246
+ "endpoint_translation_rmse_mm": 6.224776744842529,
247
+ "jerk_rmse": 3.11281418800354,
248
+ "path_rotation_rmse_deg": 2.7672295570373535,
249
+ "path_translation_rmse_mm": 3.89414119720459,
250
+ "residual_rotation_mean_deg": 0.0,
251
+ "residual_translation_mean_mm": 0.0,
252
+ "velocity_rmse": 0.040785618126392365
253
+ },
254
+ "path_score": 0.9428600311279296,
255
+ "sample_count": 520
256
+ },
257
+ "robotwin_generic": {
258
+ "group_count": 8,
259
+ "metrics": {
260
+ "acceleration_rmse": 0.12473664432764053,
261
+ "command_acceleration_rms": 0.10346533358097076,
262
+ "command_jerk_rms": 2.546473503112793,
263
+ "endpoint_rotation_rmse_deg": 3.693011522293091,
264
+ "endpoint_translation_rmse_mm": 6.344256401062012,
265
+ "jerk_rmse": 3.1005780696868896,
266
+ "path_rotation_rmse_deg": 2.3311872482299805,
267
+ "path_translation_rmse_mm": 3.962736129760742,
268
+ "residual_rotation_mean_deg": 0.0,
269
+ "residual_translation_mean_mm": 0.0,
270
+ "velocity_rmse": 0.03495059534907341
271
+ },
272
+ "path_score": 0.8625110626220703,
273
+ "sample_count": 778
274
+ }
275
+ },
276
+ "split": "validation",
277
+ "test_firewall_opened": false
278
+ }
experiments/E002/conditions/c-endpoint-5d2dff166d/run_manifest.yaml ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R001
4
+ experiment_id: E002
5
+ condition_id: c-endpoint-5d2dff166d
6
+ status: completed
7
+ started_at: '2026-08-10T11:12:08.698434+00:00'
8
+ ended_at: '2026-08-10T11:18:50.394652+00:00'
9
+ condition:
10
+ model:
11
+ action_head: endpoint
12
+ base_conditioning: none
13
+ horizon_steps: 1
14
+ trajectory_basis: none
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R001-endpoint.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples
67
+ --output-dir /home/coder/share/experiment-runs/E002/E002-R001
68
+ metrics:
69
+ validation:
70
+ loss/total: 0.21126084798840786
71
+ loss/translation: 0.04712994648100599
72
+ loss/rotation_vector: 0.04120707774795994
73
+ loss/rotation_geodesic: 0.42234639320608647
74
+ loss/joint: 0.03408178494804078
75
+ loss/gripper: 0.0029633291347672696
76
+ metric/translation_mm: 23.808171852711354
77
+ metric/rotation_deg: 12.669311751202553
78
+ metric/joint_mae_rad: 0.08809902482251357
79
+ metric/gripper_accuracy: 1.0
80
+ checkpoint:
81
+ path: /home/coder/share/experiment-runs/E002/E002-R001/best.pt
82
+ sha256: 0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da
83
+ selection_metric: loss/total
84
+ selection_split: validation
85
+ artifacts:
86
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R001/metrics.jsonl
87
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R001/evaluation_validation.json
88
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R001/condition.yaml
89
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R001/config_resolved.yaml
90
+ compute:
91
+ wall_seconds: 401.5836740859959
92
+ gpu_hours: 0.1115510205794433
93
+ peak_memory_bytes: 2885576192
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/CONDITION_CARD.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Quintic minimum-jerk CP6 selected
2
+
3
+ - Experiment / run: E002 / E002-R004
4
+ - Condition: c-minjerk-se3-cp6-af54a01c9b
5
+ - Role: validation-selected structured head
6
+ - Checkpoint SHA-256: 4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9
7
+ - Code commit: 033cb092edc8a14430f35e5c78e06323ee14655a (clean)
8
+ - Dataset: Shiki42/up-vla-precision-recovery-15k @ 77f5f306303361d602ab10759b3a139c67f2cf16
9
+ - E002 split: E001-train-only grouped re-split; train 8,807/71 groups, validation 1,298/9, fresh test 1,523/9
10
+ - Geometry: local tool frame, right SE(3) composition. Degree-5 Bernstein controls [0,0,0,q,q,q], xi(u)=(10u^3-15u^4+6u^5)q; T_cmd=T_base@Exp_SE3(xi).
11
+ - Validation: path 0.855793, endpoint 6.209 mm / 3.745 deg, command jerk 2.272
12
+ - Fresh test: path 0.835376, endpoint 5.905 mm / 3.701 deg, command jerk 2.562, GPU p95 8.214 ms
13
+ - Decision: E002 rejected; real Pi0.5 rollout not authorized
14
+
15
+ See sibling resolved config, immutable run manifest, validation report, and one-time fresh-test report.
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/best.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9
3
+ size 238735999
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/config_resolved.yaml ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
3
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
4
+ variant_id: validated-physical-recovery-15k-v1
5
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
6
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
7
+ sample_count: 15000
8
+ eligible_sample_count: 11628
9
+ eligible_group_count: 89
10
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
11
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
12
+ split_namespace: 'E002-source-group-split-v1::'
13
+ validation_fraction: 0.1
14
+ test_fraction: 0.1
15
+ geometry:
16
+ cumulative_waypoint_representation: current_tool_relative_pose
17
+ step_representation: integrated_body_twist
18
+ step_integration: right_se3_exponential
19
+ residual_curve_frame: initial_local_tool_frame
20
+ residual_composition: base_right_multiply_se3_exponential
21
+ structured_translation: se3_tangent_translation_control_points_m
22
+ structured_rotation: so3_lie_algebra_control_points_rad
23
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
24
+ residual_boundary_at_t0: identity
25
+ model:
26
+ backbone:
27
+ hidden_size: 256
28
+ fusion_layers: 4
29
+ fusion_heads: 8
30
+ fusion_mlp_ratio: 4
31
+ dropout: 0.1
32
+ vision_backend: conv
33
+ text_backend: embedding
34
+ pretrained_model_name: google/siglip-base-patch16-224
35
+ freeze_pretrained: true
36
+ image_size: 224
37
+ vision_tokens: 64
38
+ depth_tokens: 36
39
+ point_tokens: 32
40
+ point_samples: 256
41
+ max_text_tokens: 32
42
+ text_vocab_size: 32768
43
+ max_joints: 7
44
+ aloha_joints: 6
45
+ panda_joints: 7
46
+ translation_scale_m: 0.05
47
+ rotation_scale_rad: 0.5235987755982988
48
+ joint_scale_rad: 0.5
49
+ horizon_steps: 16
50
+ horizon_seconds: 0.8
51
+ decoder_layers: 4
52
+ decoder_heads: 8
53
+ decoder_mlp_ratio: 4
54
+ control_points: 6
55
+ maximum_step_translation_m: 0.01
56
+ maximum_step_rotation_rad: 0.2
57
+ maximum_residual_translation_m: 0.05
58
+ maximum_residual_rotation_rad: 0.52
59
+ action_head: minimum_jerk
60
+ base_conditioning: E002-R001-profiled-proposal
61
+ trajectory_basis: quintic_minimum_jerk_bezier_cp6
62
+ parameter_counts:
63
+ total: 20074074
64
+ trainable: 19870022
65
+ frozen: 204052
66
+ loss:
67
+ path_translation: 1.0
68
+ path_rotation: 1.0
69
+ endpoint_translation: 2.0
70
+ endpoint_rotation: 2.0
71
+ velocity: 1.0
72
+ acceleration: 0.1
73
+ jerk: 0.05
74
+ residual_magnitude: 0.02
75
+ training:
76
+ seed: 42
77
+ batch_size: 64
78
+ gradient_accumulation_steps: 1
79
+ epochs: 100
80
+ max_steps: 3000
81
+ learning_rate: 0.0003
82
+ weight_decay: 0.05
83
+ warmup_fraction: 0.05
84
+ gradient_clip_norm: 1.0
85
+ precision: bf16
86
+ workers: 8
87
+ log_every_steps: 10
88
+ validation_every_steps: 250
89
+ checkpoint_every_steps: 500
90
+ evaluation:
91
+ version: se3-spline-eval-v1
92
+ trajectory_loss:
93
+ path_translation: 1.0
94
+ path_rotation: 1.0
95
+ endpoint_translation: 2.0
96
+ endpoint_rotation: 2.0
97
+ velocity: 1.0
98
+ acceleration: 0.1
99
+ jerk: 0.05
100
+ residual_magnitude: 0.02
101
+ selection_split: validation
102
+ test_firewall_opened: false
103
+ group_bootstrap_seed: 20260810
104
+ group_bootstrap_resamples: 10000
105
+ runtime:
106
+ data_root: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples
107
+ base_proposal_root: /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
108
+ output_directory: /home/coder/share/experiment-runs/E002/E002-R004
109
+ split_counts:
110
+ train_samples: 8807
111
+ validation_samples: 1298
112
+ train_groups: 71
113
+ validation_groups: 9
114
+ overlap: false
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_fresh_test.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R004/best.pt",
8
+ "sha256": "4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R004/run_manifest.yaml",
15
+ "manifest_sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac",
16
+ "run_id": "E002-R004"
17
+ },
18
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.127791753038764,
30
+ "upper": 0.14594409606522984
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.10712421170125405,
34
+ "upper": 0.12841273289587762
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 2.2563910881678266,
38
+ "upper": 2.8369754082626764
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.017780652973387,
42
+ "upper": 4.63208531803555
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.5506986127959355,
46
+ "upper": 6.2668231990602274
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7442402338319355,
50
+ "upper": 0.9650608261426288
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 2.8284443749321833,
54
+ "upper": 3.306691540612115
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 1.9337187038527595,
58
+ "upper": 2.9215715991126165
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.4247371594111127,
62
+ "upper": 3.897781629694833
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.49182973437839084,
66
+ "upper": 0.6278564759426647
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.6452037485109435,
70
+ "upper": 0.8209328550431463
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.030529986860023603,
74
+ "upper": 0.04382059392001894
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.13714249432086945,
84
+ "command_acceleration_rms": 0.11820725351572037,
85
+ "command_jerk_rms": 2.5615224838256836,
86
+ "endpoint_rotation_rmse_deg": 3.70082688331604,
87
+ "endpoint_translation_rmse_mm": 5.904946327209473,
88
+ "jerk_rmse": 3.08172607421875,
89
+ "path_rotation_rmse_deg": 2.3470988273620605,
90
+ "path_translation_rmse_mm": 3.659562110900879,
91
+ "residual_rotation_mean_deg": 0.5654941201210022,
92
+ "residual_translation_mean_mm": 0.723316490650177,
93
+ "velocity_rmse": 0.036055486649274826
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.102456,
98
+ "p95_ms": 8.21402955,
99
+ "p99_ms": 10.41670332,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0011634723050519824,
105
+ "endpoint_rotation": 0.14235493540763855,
106
+ "endpoint_translation": 0.0025301326531916857,
107
+ "jerk": 0.0014597824774682522,
108
+ "path_rotation": 0.08016978204250336,
109
+ "path_translation": 0.0009646149701438844,
110
+ "residual_magnitude": 0.030533183366060257,
111
+ "total": 0.43282923102378845,
112
+ "velocity": 0.06112472340464592
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.1416848599910736,
116
+ "command_acceleration_rms": 0.12434567511081696,
117
+ "command_jerk_rms": 2.7526066303253174,
118
+ "endpoint_rotation_rmse_deg": 4.9904632568359375,
119
+ "endpoint_translation_rmse_mm": 6.160519123077393,
120
+ "jerk_rmse": 3.2191498279571533,
121
+ "path_rotation_rmse_deg": 3.144516944885254,
122
+ "path_translation_rmse_mm": 3.8038434982299805,
123
+ "residual_rotation_mean_deg": 0.49581214785575867,
124
+ "residual_translation_mean_mm": 0.700305700302124,
125
+ "velocity_rmse": 0.04659629613161087
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/knife::datasets/rlbench_expert_validation/samples/put_knife_in_knife_block/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.14977389574050903,
130
+ "command_acceleration_rms": 0.13863877952098846,
131
+ "command_jerk_rms": 3.1560118198394775,
132
+ "endpoint_rotation_rmse_deg": 7.083123207092285,
133
+ "endpoint_translation_rmse_mm": 6.871181488037109,
134
+ "jerk_rmse": 3.4372811317443848,
135
+ "path_rotation_rmse_deg": 4.451962471008301,
136
+ "path_translation_rmse_mm": 4.221004009246826,
137
+ "residual_rotation_mean_deg": 0.3148156702518463,
138
+ "residual_translation_mean_mm": 0.7568097114562988,
139
+ "velocity_rmse": 0.06459157913923264
140
+ },
141
+ "robotwin_generic::023_tissue-box/base1::candidate_0065": {
142
+ "acceleration_rmse": 0.13710440695285797,
143
+ "command_acceleration_rms": 0.1289472132921219,
144
+ "command_jerk_rms": 2.9070723056793213,
145
+ "endpoint_rotation_rmse_deg": 2.1072518825531006,
146
+ "endpoint_translation_rmse_mm": 6.349578380584717,
147
+ "jerk_rmse": 3.090036630630493,
148
+ "path_rotation_rmse_deg": 1.4243571758270264,
149
+ "path_translation_rmse_mm": 4.0173516273498535,
150
+ "residual_rotation_mean_deg": 0.6179807186126709,
151
+ "residual_translation_mean_mm": 1.0377143621444702,
152
+ "velocity_rmse": 0.023954471573233604
153
+ },
154
+ "robotwin_generic::049_shampoo/base1::candidate_0017": {
155
+ "acceleration_rmse": 0.11999844014644623,
156
+ "command_acceleration_rms": 0.09427337348461151,
157
+ "command_jerk_rms": 2.058751106262207,
158
+ "endpoint_rotation_rmse_deg": 3.464721918106079,
159
+ "endpoint_translation_rmse_mm": 5.774306774139404,
160
+ "jerk_rmse": 2.7728395462036133,
161
+ "path_rotation_rmse_deg": 2.1866109371185303,
162
+ "path_translation_rmse_mm": 3.5805296897888184,
163
+ "residual_rotation_mean_deg": 0.5309782028198242,
164
+ "residual_translation_mean_mm": 0.6524855494499207,
165
+ "velocity_rmse": 0.033256858587265015
166
+ },
167
+ "robotwin_generic::053_teanet/base5::candidate_0076": {
168
+ "acceleration_rmse": 0.15978120267391205,
169
+ "command_acceleration_rms": 0.1343170404434204,
170
+ "command_jerk_rms": 2.8844008445739746,
171
+ "endpoint_rotation_rmse_deg": 3.9705393314361572,
172
+ "endpoint_translation_rmse_mm": 6.632754802703857,
173
+ "jerk_rmse": 3.6311490535736084,
174
+ "path_rotation_rmse_deg": 2.536879062652588,
175
+ "path_translation_rmse_mm": 4.1681413650512695,
176
+ "residual_rotation_mean_deg": 0.7375269532203674,
177
+ "residual_translation_mean_mm": 0.8364473581314087,
178
+ "velocity_rmse": 0.03881774842739105
179
+ },
180
+ "robotwin_generic::055_small-speaker/base1::candidate_0022": {
181
+ "acceleration_rmse": 0.12982314825057983,
182
+ "command_acceleration_rms": 0.1055518388748169,
183
+ "command_jerk_rms": 2.1554129123687744,
184
+ "endpoint_rotation_rmse_deg": 3.219684600830078,
185
+ "endpoint_translation_rmse_mm": 5.317539215087891,
186
+ "jerk_rmse": 2.848005771636963,
187
+ "path_rotation_rmse_deg": 2.036024570465088,
188
+ "path_translation_rmse_mm": 3.2720251083374023,
189
+ "residual_rotation_mean_deg": 0.5957156419754028,
190
+ "residual_translation_mean_mm": 0.6151946783065796,
191
+ "velocity_rmse": 0.03162113577127457
192
+ },
193
+ "robotwin_generic::073_rubikscube/base0::candidate_0030": {
194
+ "acceleration_rmse": 0.14153644442558289,
195
+ "command_acceleration_rms": 0.1246294379234314,
196
+ "command_jerk_rms": 2.7487876415252686,
197
+ "endpoint_rotation_rmse_deg": 3.0712850093841553,
198
+ "endpoint_translation_rmse_mm": 5.070908546447754,
199
+ "jerk_rmse": 3.1885945796966553,
200
+ "path_rotation_rmse_deg": 1.941308617591858,
201
+ "path_translation_rmse_mm": 3.116575241088867,
202
+ "residual_rotation_mean_deg": 0.5864089131355286,
203
+ "residual_translation_mean_mm": 0.5934191942214966,
204
+ "velocity_rmse": 0.03096099942922592
205
+ },
206
+ "robotwin_generic::077_phone/base1::candidate_0087": {
207
+ "acceleration_rmse": 0.1446823626756668,
208
+ "command_acceleration_rms": 0.12868784368038177,
209
+ "command_jerk_rms": 2.852937936782837,
210
+ "endpoint_rotation_rmse_deg": 3.3648030757904053,
211
+ "endpoint_translation_rmse_mm": 5.805663585662842,
212
+ "jerk_rmse": 3.2859463691711426,
213
+ "path_rotation_rmse_deg": 2.1426286697387695,
214
+ "path_translation_rmse_mm": 3.5831217765808105,
215
+ "residual_rotation_mean_deg": 0.5478863716125488,
216
+ "residual_translation_mean_mm": 0.5915248394012451,
217
+ "velocity_rmse": 0.03348454833030701
218
+ },
219
+ "robotwin_generic::107_soap/base0::candidate_0052": {
220
+ "acceleration_rmse": 0.14021557569503784,
221
+ "command_acceleration_rms": 0.118662990629673,
222
+ "command_jerk_rms": 2.590258836746216,
223
+ "endpoint_rotation_rmse_deg": 3.489109754562378,
224
+ "endpoint_translation_rmse_mm": 5.66431188583374,
225
+ "jerk_rmse": 3.1714670658111572,
226
+ "path_rotation_rmse_deg": 2.2048773765563965,
227
+ "path_translation_rmse_mm": 3.510601758956909,
228
+ "residual_rotation_mean_deg": 0.60318922996521,
229
+ "residual_translation_mean_mm": 0.6841212511062622,
230
+ "velocity_rmse": 0.034124862402677536
231
+ },
232
+ "robotwin_generic::109_hydrating-oil/base1::candidate_0107": {
233
+ "acceleration_rmse": 0.11136701703071594,
234
+ "command_acceleration_rms": 0.09015678614377975,
235
+ "command_jerk_rms": 1.7000700235366821,
236
+ "endpoint_rotation_rmse_deg": 3.5369248390197754,
237
+ "endpoint_translation_rmse_mm": 5.658270835876465,
238
+ "jerk_rmse": 2.310215711593628,
239
+ "path_rotation_rmse_deg": 2.199239730834961,
240
+ "path_translation_rmse_mm": 3.4667046070098877,
241
+ "residual_rotation_mean_deg": 0.5549455881118774,
242
+ "residual_translation_mean_mm": 0.742131769657135,
243
+ "velocity_rmse": 0.033687133342027664
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8353759765625,
250
+ "micro": 1.0092877388000487
251
+ },
252
+ "run_id": "E002-R004",
253
+ "sample_count": 1523,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.14977389574050903,
260
+ "command_acceleration_rms": 0.13863877952098846,
261
+ "command_jerk_rms": 3.1560118198394775,
262
+ "endpoint_rotation_rmse_deg": 7.083123207092285,
263
+ "endpoint_translation_rmse_mm": 6.871181488037109,
264
+ "jerk_rmse": 3.4372811317443848,
265
+ "path_rotation_rmse_deg": 4.451962471008301,
266
+ "path_translation_rmse_mm": 4.221004009246826,
267
+ "residual_rotation_mean_deg": 0.3148156702518463,
268
+ "residual_translation_mean_mm": 0.7568097114562988,
269
+ "velocity_rmse": 0.06459157913923264
270
+ },
271
+ "path_score": 1.3124928951263426,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.13730370998382568,
278
+ "command_acceleration_rms": 0.11624578386545181,
279
+ "command_jerk_rms": 2.518153190612793,
280
+ "endpoint_rotation_rmse_deg": 3.4359488487243652,
281
+ "endpoint_translation_rmse_mm": 5.757655620574951,
282
+ "jerk_rmse": 3.100025177001953,
283
+ "path_rotation_rmse_deg": 2.1768791675567627,
284
+ "path_translation_rmse_mm": 3.568424701690674,
285
+ "residual_rotation_mean_deg": 0.5896488428115845,
286
+ "residual_translation_mean_mm": 0.6710115075111389,
287
+ "velocity_rmse": 0.03367314860224724
288
+ },
289
+ "path_score": 0.79221830368042,
290
+ "sample_count": 1003
291
+ }
292
+ },
293
+ "split": "test",
294
+ "test_firewall_opened": true
295
+ }
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/evaluation_validation.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R004/best.pt",
8
+ "sha256": "4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R004/run_manifest.yaml",
15
+ "manifest_sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac",
16
+ "run_id": "E002-R004"
17
+ },
18
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.10308757486442725,
30
+ "upper": 0.1397645959837569
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.08936153536455499,
34
+ "upper": 0.11839342810627487
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 1.719706772930092,
38
+ "upper": 2.6594989677270253
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.205244329240587,
42
+ "upper": 4.427248266008165
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.687558265527089,
46
+ "upper": 6.8257540742556255
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7683746231926811,
50
+ "upper": 0.9666738802856869
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 2.1001228921943245,
54
+ "upper": 3.2242833753426865
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 2.028298435277409,
58
+ "upper": 2.7444795823759502
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.546643653843138,
62
+ "upper": 4.226622504658169
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.5349525012903743,
66
+ "upper": 0.5974235445261001
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.7040380966332224,
70
+ "upper": 0.9369111968411339
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.03159253713157442,
74
+ "upper": 0.041467026931544144
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.12487276643514633,
84
+ "command_acceleration_rms": 0.10622868686914444,
85
+ "command_jerk_rms": 2.2715606689453125,
86
+ "endpoint_rotation_rmse_deg": 3.7452497482299805,
87
+ "endpoint_translation_rmse_mm": 6.209352493286133,
88
+ "jerk_rmse": 2.7772092819213867,
89
+ "path_rotation_rmse_deg": 2.348088264465332,
90
+ "path_translation_rmse_mm": 3.861748695373535,
91
+ "residual_rotation_mean_deg": 0.5649371147155762,
92
+ "residual_translation_mean_mm": 0.8120338320732117,
93
+ "velocity_rmse": 0.0359807126224041
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.1037045,
98
+ "p95_ms": 8.980220799999982,
99
+ "p99_ms": 10.40966553,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0013029025867581367,
105
+ "endpoint_rotation": 0.10964572429656982,
106
+ "endpoint_translation": 0.0024704034440219402,
107
+ "jerk": 0.0016885013319551945,
108
+ "path_rotation": 0.06242626905441284,
109
+ "path_translation": 0.0009601161582395434,
110
+ "residual_magnitude": 0.03481997549533844,
111
+ "total": 0.34016457200050354,
112
+ "velocity": 0.05163482204079628
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.1349928230047226,
116
+ "command_acceleration_rms": 0.11819975078105927,
117
+ "command_jerk_rms": 2.664361000061035,
118
+ "endpoint_rotation_rmse_deg": 3.799849510192871,
119
+ "endpoint_translation_rmse_mm": 6.087368965148926,
120
+ "jerk_rmse": 3.123729705810547,
121
+ "path_rotation_rmse_deg": 2.3968474864959717,
122
+ "path_translation_rmse_mm": 3.7949631214141846,
123
+ "residual_rotation_mean_deg": 0.5683494806289673,
124
+ "residual_translation_mean_mm": 0.7937496900558472,
125
+ "velocity_rmse": 0.03653155639767647
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.1344926357269287,
130
+ "command_acceleration_rms": 0.12197227776050568,
131
+ "command_jerk_rms": 2.801678419113159,
132
+ "endpoint_rotation_rmse_deg": 4.230731964111328,
133
+ "endpoint_translation_rmse_mm": 6.085890293121338,
134
+ "jerk_rmse": 3.1305017471313477,
135
+ "path_rotation_rmse_deg": 2.674309253692627,
136
+ "path_translation_rmse_mm": 3.8025522232055664,
137
+ "residual_rotation_mean_deg": 0.5871337652206421,
138
+ "residual_translation_mean_mm": 0.9068980813026428,
139
+ "velocity_rmse": 0.040271393954753876
140
+ },
141
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
142
+ "acceleration_rmse": 0.1561744213104248,
143
+ "command_acceleration_rms": 0.1311158388853073,
144
+ "command_jerk_rms": 3.008500099182129,
145
+ "endpoint_rotation_rmse_deg": 3.859532117843628,
146
+ "endpoint_translation_rmse_mm": 7.020136833190918,
147
+ "jerk_rmse": 3.663717269897461,
148
+ "path_rotation_rmse_deg": 2.4413540363311768,
149
+ "path_translation_rmse_mm": 4.408639907836914,
150
+ "residual_rotation_mean_deg": 0.5009645819664001,
151
+ "residual_translation_mean_mm": 0.7318453788757324,
152
+ "velocity_rmse": 0.03755943849682808
153
+ },
154
+ "robotwin_generic::050_bell/base0::candidate_0018": {
155
+ "acceleration_rmse": 0.12712980806827545,
156
+ "command_acceleration_rms": 0.11159520596265793,
157
+ "command_jerk_rms": 2.3976645469665527,
158
+ "endpoint_rotation_rmse_deg": 3.7091081142425537,
159
+ "endpoint_translation_rmse_mm": 5.565788745880127,
160
+ "jerk_rmse": 2.9391865730285645,
161
+ "path_rotation_rmse_deg": 2.3155221939086914,
162
+ "path_translation_rmse_mm": 3.4275779724121094,
163
+ "residual_rotation_mean_deg": 0.5825850367546082,
164
+ "residual_translation_mean_mm": 0.713529646396637,
165
+ "velocity_rmse": 0.03522006422281265
166
+ },
167
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
168
+ "acceleration_rmse": 0.04633811488747597,
169
+ "command_acceleration_rms": 0.046336181461811066,
170
+ "command_jerk_rms": 0.2913300395011902,
171
+ "endpoint_rotation_rmse_deg": 6.0872697830200195,
172
+ "endpoint_translation_rmse_mm": 8.130516052246094,
173
+ "jerk_rmse": 0.2913312017917633,
174
+ "path_rotation_rmse_deg": 3.6939315795898438,
175
+ "path_translation_rmse_mm": 4.995799541473389,
176
+ "residual_rotation_mean_deg": 0.5199920535087585,
177
+ "residual_translation_mean_mm": 1.134708046913147,
178
+ "velocity_rmse": 0.05466964840888977
179
+ },
180
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
181
+ "acceleration_rmse": 0.13279545307159424,
182
+ "command_acceleration_rms": 0.11015186458826065,
183
+ "command_jerk_rms": 2.5362021923065186,
184
+ "endpoint_rotation_rmse_deg": 2.6013705730438232,
185
+ "endpoint_translation_rmse_mm": 6.034164905548096,
186
+ "jerk_rmse": 3.1239142417907715,
187
+ "path_rotation_rmse_deg": 1.6837923526763916,
188
+ "path_translation_rmse_mm": 3.8183398246765137,
189
+ "residual_rotation_mean_deg": 0.5236051082611084,
190
+ "residual_translation_mean_mm": 0.7148460149765015,
191
+ "velocity_rmse": 0.026725247502326965
192
+ },
193
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
194
+ "acceleration_rmse": 0.1420787125825882,
195
+ "command_acceleration_rms": 0.09741360694169998,
196
+ "command_jerk_rms": 2.0236129760742188,
197
+ "endpoint_rotation_rmse_deg": 2.8853201866149902,
198
+ "endpoint_translation_rmse_mm": 4.953864574432373,
199
+ "jerk_rmse": 3.1655778884887695,
200
+ "path_rotation_rmse_deg": 1.8346418142318726,
201
+ "path_translation_rmse_mm": 3.1223437786102295,
202
+ "residual_rotation_mean_deg": 0.6529508829116821,
203
+ "residual_translation_mean_mm": 1.0946613550186157,
204
+ "velocity_rmse": 0.02910573035478592
205
+ },
206
+ "robotwin_generic::071_can/base1::candidate_0084": {
207
+ "acceleration_rmse": 0.12845297157764435,
208
+ "command_acceleration_rms": 0.1191152036190033,
209
+ "command_jerk_rms": 2.6901283264160156,
210
+ "endpoint_rotation_rmse_deg": 3.2490086555480957,
211
+ "endpoint_translation_rmse_mm": 6.103634834289551,
212
+ "jerk_rmse": 2.9270081520080566,
213
+ "path_rotation_rmse_deg": 2.0134360790252686,
214
+ "path_translation_rmse_mm": 3.7747769355773926,
215
+ "residual_rotation_mean_deg": 0.5559494495391846,
216
+ "residual_translation_mean_mm": 0.6914225816726685,
217
+ "velocity_rmse": 0.031804319471120834
218
+ },
219
+ "robotwin_generic::107_soap/base1::candidate_0106": {
220
+ "acceleration_rmse": 0.13067100942134857,
221
+ "command_acceleration_rms": 0.10602162033319473,
222
+ "command_jerk_rms": 2.293989419937134,
223
+ "endpoint_rotation_rmse_deg": 3.491788625717163,
224
+ "endpoint_translation_rmse_mm": 5.5043182373046875,
225
+ "jerk_rmse": 2.970918893814087,
226
+ "path_rotation_rmse_deg": 2.2204670906066895,
227
+ "path_translation_rmse_mm": 3.416995048522949,
228
+ "residual_rotation_mean_deg": 0.5371366143226624,
229
+ "residual_translation_mean_mm": 0.600003719329834,
230
+ "velocity_rmse": 0.03396544232964516
231
+ },
232
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
233
+ "acceleration_rmse": 0.12572182714939117,
234
+ "command_acceleration_rms": 0.11233633756637573,
235
+ "command_jerk_rms": 2.4009382724761963,
236
+ "endpoint_rotation_rmse_deg": 3.593118906021118,
237
+ "endpoint_translation_rmse_mm": 6.4858598709106445,
238
+ "jerk_rmse": 2.7827272415161133,
239
+ "path_rotation_rmse_deg": 2.255338430404663,
240
+ "path_translation_rmse_mm": 3.9887139797210693,
241
+ "residual_rotation_mean_deg": 0.6241169571876526,
242
+ "residual_translation_mean_mm": 0.7203899025917053,
243
+ "velocity_rmse": 0.034505102783441544
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8557925224304199,
250
+ "micro": 0.8588658094406127
251
+ },
252
+ "run_id": "E002-R004",
253
+ "sample_count": 1298,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.1344926357269287,
260
+ "command_acceleration_rms": 0.12197227776050568,
261
+ "command_jerk_rms": 2.801678419113159,
262
+ "endpoint_rotation_rmse_deg": 4.230731964111328,
263
+ "endpoint_translation_rmse_mm": 6.085890293121338,
264
+ "jerk_rmse": 3.1305017471313477,
265
+ "path_rotation_rmse_deg": 2.674309253692627,
266
+ "path_translation_rmse_mm": 3.8025522232055664,
267
+ "residual_rotation_mean_deg": 0.5871337652206421,
268
+ "residual_translation_mean_mm": 0.9068980813026428,
269
+ "velocity_rmse": 0.040271393954753876
270
+ },
271
+ "path_score": 0.9151170730590821,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.135326087474823,
278
+ "command_acceleration_rms": 0.11560964584350586,
279
+ "command_jerk_rms": 2.568490743637085,
280
+ "endpoint_rotation_rmse_deg": 3.482255220413208,
281
+ "endpoint_translation_rmse_mm": 6.0883564949035645,
282
+ "jerk_rmse": 3.119194984436035,
283
+ "path_rotation_rmse_deg": 2.191901683807373,
284
+ "path_translation_rmse_mm": 3.789881944656372,
285
+ "residual_rotation_mean_deg": 0.5557944178581238,
286
+ "residual_translation_mean_mm": 0.7181235551834106,
287
+ "velocity_rmse": 0.03380199894309044
288
+ },
289
+ "path_score": 0.8173685312271117,
290
+ "sample_count": 778
291
+ }
292
+ },
293
+ "split": "validation",
294
+ "test_firewall_opened": false
295
+ }
experiments/E002/conditions/c-minjerk-se3-cp6-af54a01c9b/run_manifest.yaml ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R004
4
+ experiment_id: E002
5
+ condition_id: c-minjerk-se3-cp6-af54a01c9b
6
+ status: completed
7
+ started_at: '2026-08-10T11:42:37.552494+00:00'
8
+ ended_at: '2026-08-10T11:47:36.515256+00:00'
9
+ condition:
10
+ model:
11
+ action_head: minimum_jerk
12
+ base_conditioning: E002-R001-profiled-proposal
13
+ horizon_steps: 16
14
+ trajectory_basis: quintic_minimum_jerk_bezier_cp6
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R004-minjerk-se3-cp6.yaml --data-root
67
+ /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir
68
+ /home/coder/share/experiment-runs/E002/E002-R004 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
69
+ metrics:
70
+ validation:
71
+ loss/total: 0.340089762706786
72
+ loss/path_translation: 0.0009601259408934825
73
+ loss/path_rotation: 0.062407021208426616
74
+ loss/endpoint_translation: 0.002470431426390776
75
+ loss/endpoint_rotation: 0.10961773915816161
76
+ loss/velocity: 0.051635161111094
77
+ loss/acceleration: 0.0013029350546147037
78
+ loss/jerk: 0.0016885478681280565
79
+ loss/residual_magnitude: 0.034819636315979834
80
+ metric/path_translation_rmse_mm: 3.7914421139586323
81
+ metric/path_rotation_rmse_deg: 2.3811969316244492
82
+ metric/endpoint_translation_rmse_mm: 6.081028670118476
83
+ metric/endpoint_rotation_rmse_deg: 3.774994674559183
84
+ metric/velocity_rmse: 0.036343481690150746
85
+ metric/acceleration_rmse: 0.13480377996280124
86
+ metric/jerk_rmse: 3.113585577540111
87
+ metric/command_acceleration_rms: 0.11796897273217952
88
+ metric/command_jerk_rms: 2.653614019576133
89
+ metric/residual_translation_mean_mm: 0.7937416224891122
90
+ metric/residual_rotation_mean_deg: 0.5683441197008858
91
+ checkpoint:
92
+ path: /home/coder/share/experiment-runs/E002/E002-R004/best.pt
93
+ sha256: 4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9
94
+ selection_metric: loss/total
95
+ selection_split: validation
96
+ artifacts:
97
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R004/metrics.jsonl
98
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R004/evaluation_validation.json
99
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R004/condition.yaml
100
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R004/config_resolved.yaml
101
+ compute:
102
+ wall_seconds: 298.82912589100306
103
+ gpu_hours: 0.08300809052527863
104
+ peak_memory_bytes: 2951936512
experiments/E002/conditions/c-residual-h16-f186e90df2/CONDITION_CARD.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Residual H16 baseline
2
+
3
+ - Experiment / run: E002 / E002-R002
4
+ - Condition: c-residual-h16-f186e90df2
5
+ - Role: baseline
6
+ - Checkpoint SHA-256: c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9
7
+ - Code commit: 033cb092edc8a14430f35e5c78e06323ee14655a (clean)
8
+ - Dataset: Shiki42/up-vla-precision-recovery-15k @ 77f5f306303361d602ab10759b3a139c67f2cf16
9
+ - E002 split: E001-train-only grouped re-split; train 8,807/71 groups, validation 1,298/9, fresh test 1,523/9
10
+ - Geometry: local tool frame, right SE(3) composition. Independent 16-waypoint residuals.
11
+ - Validation: path 0.856968, endpoint 6.062 mm / 3.858 deg, command jerk 3.137
12
+ - Fresh test: path 0.801884, endpoint 5.905 mm / 3.462 deg, command jerk 3.143, GPU p95 8.237 ms
13
+ - Decision: E002 rejected; real Pi0.5 rollout not authorized
14
+
15
+ See sibling resolved config, immutable run manifest, validation report, and one-time fresh-test report.
experiments/E002/conditions/c-residual-h16-f186e90df2/best.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9
3
+ size 238782079
experiments/E002/conditions/c-residual-h16-f186e90df2/config_resolved.yaml ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
3
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
4
+ variant_id: validated-physical-recovery-15k-v1
5
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
6
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
7
+ sample_count: 15000
8
+ eligible_sample_count: 11628
9
+ eligible_group_count: 89
10
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
11
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
12
+ split_namespace: 'E002-source-group-split-v1::'
13
+ validation_fraction: 0.1
14
+ test_fraction: 0.1
15
+ geometry:
16
+ cumulative_waypoint_representation: current_tool_relative_pose
17
+ step_representation: integrated_body_twist
18
+ step_integration: right_se3_exponential
19
+ residual_curve_frame: initial_local_tool_frame
20
+ residual_composition: base_right_multiply_se3_exponential
21
+ structured_translation: se3_tangent_translation_control_points_m
22
+ structured_rotation: so3_lie_algebra_control_points_rad
23
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
24
+ residual_boundary_at_t0: identity
25
+ model:
26
+ backbone:
27
+ hidden_size: 256
28
+ fusion_layers: 4
29
+ fusion_heads: 8
30
+ fusion_mlp_ratio: 4
31
+ dropout: 0.1
32
+ vision_backend: conv
33
+ text_backend: embedding
34
+ pretrained_model_name: google/siglip-base-patch16-224
35
+ freeze_pretrained: true
36
+ image_size: 224
37
+ vision_tokens: 64
38
+ depth_tokens: 36
39
+ point_tokens: 32
40
+ point_samples: 256
41
+ max_text_tokens: 32
42
+ text_vocab_size: 32768
43
+ max_joints: 7
44
+ aloha_joints: 6
45
+ panda_joints: 7
46
+ translation_scale_m: 0.05
47
+ rotation_scale_rad: 0.5235987755982988
48
+ joint_scale_rad: 0.5
49
+ horizon_steps: 16
50
+ horizon_seconds: 0.8
51
+ decoder_layers: 4
52
+ decoder_heads: 8
53
+ decoder_mlp_ratio: 4
54
+ control_points: 6
55
+ maximum_step_translation_m: 0.01
56
+ maximum_step_rotation_rad: 0.2
57
+ maximum_residual_translation_m: 0.05
58
+ maximum_residual_rotation_rad: 0.52
59
+ action_head: residual_chunk
60
+ base_conditioning: E002-R001-profiled-proposal
61
+ trajectory_basis: independent_h16
62
+ parameter_counts:
63
+ total: 20077914
64
+ trainable: 19873862
65
+ frozen: 204052
66
+ loss:
67
+ path_translation: 1.0
68
+ path_rotation: 1.0
69
+ endpoint_translation: 2.0
70
+ endpoint_rotation: 2.0
71
+ velocity: 1.0
72
+ acceleration: 0.1
73
+ jerk: 0.05
74
+ residual_magnitude: 0.02
75
+ training:
76
+ seed: 42
77
+ batch_size: 64
78
+ gradient_accumulation_steps: 1
79
+ epochs: 100
80
+ max_steps: 3000
81
+ learning_rate: 0.0003
82
+ weight_decay: 0.05
83
+ warmup_fraction: 0.05
84
+ gradient_clip_norm: 1.0
85
+ precision: bf16
86
+ workers: 8
87
+ log_every_steps: 10
88
+ validation_every_steps: 250
89
+ checkpoint_every_steps: 500
90
+ evaluation:
91
+ version: se3-spline-eval-v1
92
+ trajectory_loss:
93
+ path_translation: 1.0
94
+ path_rotation: 1.0
95
+ endpoint_translation: 2.0
96
+ endpoint_rotation: 2.0
97
+ velocity: 1.0
98
+ acceleration: 0.1
99
+ jerk: 0.05
100
+ residual_magnitude: 0.02
101
+ selection_split: validation
102
+ test_firewall_opened: false
103
+ group_bootstrap_seed: 20260810
104
+ group_bootstrap_resamples: 10000
105
+ runtime:
106
+ data_root: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples
107
+ base_proposal_root: /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
108
+ output_directory: /home/coder/share/experiment-runs/E002/E002-R002
109
+ split_counts:
110
+ train_samples: 8807
111
+ validation_samples: 1298
112
+ train_groups: 71
113
+ validation_groups: 9
114
+ overlap: false
experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_fresh_test.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R002/best.pt",
8
+ "sha256": "c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-residual-h16-f186e90df2",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R002/run_manifest.yaml",
15
+ "manifest_sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73",
16
+ "run_id": "E002-R002"
17
+ },
18
+ "condition_id": "c-residual-h16-f186e90df2",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.13074185450871786,
30
+ "upper": 0.16456541187233395
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.11084854687667554,
34
+ "upper": 0.1475676911158694
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 2.763100173738268,
38
+ "upper": 3.5584357129202946
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 2.8186066928837032,
42
+ "upper": 4.241162343819936
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.567114898893569,
46
+ "upper": 6.269226944446563
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7137930199835036,
50
+ "upper": 0.9175362461805343
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 3.2203344961007434,
54
+ "upper": 3.95932028359837
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 1.7755660480923123,
58
+ "upper": 2.6797972198989655
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.434561617506875,
62
+ "upper": 3.892301390568415
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.6014344857798682,
66
+ "upper": 0.8241439182725218
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.5710233522786035,
70
+ "upper": 0.6960068250695863
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.027650315987153184,
74
+ "upper": 0.04000451112062567
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.14657944440841675,
84
+ "command_acceleration_rms": 0.1282518208026886,
85
+ "command_jerk_rms": 3.143152952194214,
86
+ "endpoint_rotation_rmse_deg": 3.4617953300476074,
87
+ "endpoint_translation_rmse_mm": 5.905381202697754,
88
+ "jerk_rmse": 3.5692028999328613,
89
+ "path_rotation_rmse_deg": 2.180569648742676,
90
+ "path_translation_rmse_mm": 3.6577022075653076,
91
+ "residual_rotation_mean_deg": 0.7084853053092957,
92
+ "residual_translation_mean_mm": 0.6295439600944519,
93
+ "velocity_rmse": 0.03316418454051018
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.124232,
98
+ "p95_ms": 8.236784049999999,
99
+ "p99_ms": 10.43429551,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0014884433476254344,
105
+ "endpoint_rotation": 0.129348024725914,
106
+ "endpoint_translation": 0.002536413259804249,
107
+ "jerk": 0.0020586922764778137,
108
+ "path_rotation": 0.07295631617307663,
109
+ "path_translation": 0.0009676080080680549,
110
+ "residual_magnitude": 0.033414337784051895,
111
+ "total": 0.3949826955795288,
112
+ "velocity": 0.056369852274656296
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.16159752011299133,
116
+ "command_acceleration_rms": 0.1462641954421997,
117
+ "command_jerk_rms": 3.4728081226348877,
118
+ "endpoint_rotation_rmse_deg": 4.51934814453125,
119
+ "endpoint_translation_rmse_mm": 6.1681599617004395,
120
+ "jerk_rmse": 3.84936261177063,
121
+ "path_rotation_rmse_deg": 2.86142635345459,
122
+ "path_translation_rmse_mm": 3.8097403049468994,
123
+ "residual_rotation_mean_deg": 0.6457820534706116,
124
+ "residual_translation_mean_mm": 0.5944134593009949,
125
+ "velocity_rmse": 0.04231090471148491
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/knife::datasets/rlbench_expert_validation/samples/put_knife_in_knife_block/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.18965737521648407,
130
+ "command_acceleration_rms": 0.1810983568429947,
131
+ "command_jerk_rms": 4.145353317260742,
132
+ "endpoint_rotation_rmse_deg": 6.237173080444336,
133
+ "endpoint_translation_rmse_mm": 6.932553291320801,
134
+ "jerk_rmse": 4.3626604080200195,
135
+ "path_rotation_rmse_deg": 3.963348865509033,
136
+ "path_translation_rmse_mm": 4.267070770263672,
137
+ "residual_rotation_mean_deg": 0.46613171696662903,
138
+ "residual_translation_mean_mm": 0.566665530204773,
139
+ "velocity_rmse": 0.0575411431491375
140
+ },
141
+ "robotwin_generic::023_tissue-box/base1::candidate_0065": {
142
+ "acceleration_rmse": 0.12910345196723938,
143
+ "command_acceleration_rms": 0.1204143762588501,
144
+ "command_jerk_rms": 3.129164695739746,
145
+ "endpoint_rotation_rmse_deg": 1.7799073457717896,
146
+ "endpoint_translation_rmse_mm": 6.32249641418457,
147
+ "jerk_rmse": 3.261528968811035,
148
+ "path_rotation_rmse_deg": 1.1387267112731934,
149
+ "path_translation_rmse_mm": 3.9910120964050293,
150
+ "residual_rotation_mean_deg": 0.6829317212104797,
151
+ "residual_translation_mean_mm": 0.7053245902061462,
152
+ "velocity_rmse": 0.01903185434639454
153
+ },
154
+ "robotwin_generic::049_shampoo/base1::candidate_0017": {
155
+ "acceleration_rmse": 0.12449102848768234,
156
+ "command_acceleration_rms": 0.0981483906507492,
157
+ "command_jerk_rms": 2.4942703247070312,
158
+ "endpoint_rotation_rmse_deg": 3.3874828815460205,
159
+ "endpoint_translation_rmse_mm": 5.680295944213867,
160
+ "jerk_rmse": 3.1067166328430176,
161
+ "path_rotation_rmse_deg": 2.124467611312866,
162
+ "path_translation_rmse_mm": 3.519031047821045,
163
+ "residual_rotation_mean_deg": 0.6844011545181274,
164
+ "residual_translation_mean_mm": 0.5747062563896179,
165
+ "velocity_rmse": 0.032143741846084595
166
+ },
167
+ "robotwin_generic::053_teanet/base5::candidate_0076": {
168
+ "acceleration_rmse": 0.19514064490795135,
169
+ "command_acceleration_rms": 0.1735788881778717,
170
+ "command_jerk_rms": 4.18446683883667,
171
+ "endpoint_rotation_rmse_deg": 3.9870221614837646,
172
+ "endpoint_translation_rmse_mm": 6.59793758392334,
173
+ "jerk_rmse": 4.707324028015137,
174
+ "path_rotation_rmse_deg": 2.512798309326172,
175
+ "path_translation_rmse_mm": 4.137022495269775,
176
+ "residual_rotation_mean_deg": 1.015197992324829,
177
+ "residual_translation_mean_mm": 0.6522755026817322,
178
+ "velocity_rmse": 0.038124632090330124
179
+ },
180
+ "robotwin_generic::055_small-speaker/base1::candidate_0022": {
181
+ "acceleration_rmse": 0.13783422112464905,
182
+ "command_acceleration_rms": 0.11412148177623749,
183
+ "command_jerk_rms": 2.7844250202178955,
184
+ "endpoint_rotation_rmse_deg": 3.0998964309692383,
185
+ "endpoint_translation_rmse_mm": 5.337655544281006,
186
+ "jerk_rmse": 3.355896472930908,
187
+ "path_rotation_rmse_deg": 1.9447523355484009,
188
+ "path_translation_rmse_mm": 3.289301872253418,
189
+ "residual_rotation_mean_deg": 0.5801345705986023,
190
+ "residual_translation_mean_mm": 0.5789918303489685,
191
+ "velocity_rmse": 0.02984410524368286
192
+ },
193
+ "robotwin_generic::073_rubikscube/base0::candidate_0030": {
194
+ "acceleration_rmse": 0.14262448251247406,
195
+ "command_acceleration_rms": 0.1255357414484024,
196
+ "command_jerk_rms": 3.1200077533721924,
197
+ "endpoint_rotation_rmse_deg": 3.0360608100891113,
198
+ "endpoint_translation_rmse_mm": 5.193727970123291,
199
+ "jerk_rmse": 3.5244839191436768,
200
+ "path_rotation_rmse_deg": 1.8893157243728638,
201
+ "path_translation_rmse_mm": 3.1941301822662354,
202
+ "residual_rotation_mean_deg": 0.8960033059120178,
203
+ "residual_translation_mean_mm": 0.5647693872451782,
204
+ "velocity_rmse": 0.029465213418006897
205
+ },
206
+ "robotwin_generic::077_phone/base1::candidate_0087": {
207
+ "acceleration_rmse": 0.14364629983901978,
208
+ "command_acceleration_rms": 0.12714022397994995,
209
+ "command_jerk_rms": 3.114225387573242,
210
+ "endpoint_rotation_rmse_deg": 3.2559661865234375,
211
+ "endpoint_translation_rmse_mm": 5.799946308135986,
212
+ "jerk_rmse": 3.5136070251464844,
213
+ "path_rotation_rmse_deg": 2.0622518062591553,
214
+ "path_translation_rmse_mm": 3.586200714111328,
215
+ "residual_rotation_mean_deg": 0.5341243147850037,
216
+ "residual_translation_mean_mm": 0.4976072311401367,
217
+ "velocity_rmse": 0.03169773519039154
218
+ },
219
+ "robotwin_generic::107_soap/base0::candidate_0052": {
220
+ "acceleration_rmse": 0.14373904466629028,
221
+ "command_acceleration_rms": 0.12301642447710037,
222
+ "command_jerk_rms": 3.0577945709228516,
223
+ "endpoint_rotation_rmse_deg": 3.1052424907684326,
224
+ "endpoint_translation_rmse_mm": 5.576314926147461,
225
+ "jerk_rmse": 3.553825855255127,
226
+ "path_rotation_rmse_deg": 1.958408236503601,
227
+ "path_translation_rmse_mm": 3.444178581237793,
228
+ "residual_rotation_mean_deg": 0.8805226683616638,
229
+ "residual_translation_mean_mm": 0.6912943720817566,
230
+ "velocity_rmse": 0.02998124249279499
231
+ },
232
+ "robotwin_generic::109_hydrating-oil/base1::candidate_0107": {
233
+ "acceleration_rmse": 0.11297843605279922,
234
+ "command_acceleration_rms": 0.09121248871088028,
235
+ "command_jerk_rms": 2.258666753768921,
236
+ "endpoint_rotation_rmse_deg": 3.2674081325531006,
237
+ "endpoint_translation_rmse_mm": 5.707505226135254,
238
+ "jerk_rmse": 2.736780881881714,
239
+ "path_rotation_rmse_deg": 2.0310559272766113,
240
+ "path_translation_rmse_mm": 3.4913723468780518,
241
+ "residual_rotation_mean_deg": 0.636920154094696,
242
+ "residual_translation_mean_mm": 0.8342603445053101,
243
+ "velocity_rmse": 0.030647991225123405
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8018841505050659,
250
+ "micro": 0.9532593011856079
251
+ },
252
+ "run_id": "E002-R002",
253
+ "sample_count": 1523,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.18965737521648407,
260
+ "command_acceleration_rms": 0.1810983568429947,
261
+ "command_jerk_rms": 4.145353317260742,
262
+ "endpoint_rotation_rmse_deg": 6.237173080444336,
263
+ "endpoint_translation_rmse_mm": 6.932553291320801,
264
+ "jerk_rmse": 4.3626604080200195,
265
+ "path_rotation_rmse_deg": 3.963348865509033,
266
+ "path_translation_rmse_mm": 4.267070770263672,
267
+ "residual_rotation_mean_deg": 0.46613171696662903,
268
+ "residual_translation_mean_mm": 0.566665530204773,
269
+ "velocity_rmse": 0.0575411431491375
270
+ },
271
+ "path_score": 1.2193768501281739,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.14492717385292053,
278
+ "command_acceleration_rms": 0.12442337721586227,
279
+ "command_jerk_rms": 3.06661319732666,
280
+ "endpoint_rotation_rmse_deg": 3.2931413650512695,
281
+ "endpoint_translation_rmse_mm": 5.731878757476807,
282
+ "jerk_rmse": 3.5541861057281494,
283
+ "path_rotation_rmse_deg": 2.070955991744995,
284
+ "path_translation_rmse_mm": 3.5495212078094482,
285
+ "residual_rotation_mean_deg": 0.7389209270477295,
286
+ "residual_translation_mean_mm": 0.608799159526825,
287
+ "velocity_rmse": 0.03165086358785629
288
+ },
289
+ "path_score": 0.7691433191299439,
290
+ "sample_count": 1003
291
+ }
292
+ },
293
+ "split": "test",
294
+ "test_firewall_opened": true
295
+ }
experiments/E002/conditions/c-residual-h16-f186e90df2/evaluation_validation.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R002/best.pt",
8
+ "sha256": "c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-residual-h16-f186e90df2",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R002/run_manifest.yaml",
15
+ "manifest_sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73",
16
+ "run_id": "E002-R002"
17
+ },
18
+ "condition_id": "c-residual-h16-f186e90df2",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.12913274992671278,
30
+ "upper": 0.15617000725534227
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.11790996123519208,
34
+ "upper": 0.13553185007638402
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 2.875136958228217,
38
+ "upper": 3.3160451802942488
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.29721736907959,
42
+ "upper": 4.666348218917846
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.61740255355835,
46
+ "upper": 6.495758643415239
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7692517068650987,
50
+ "upper": 0.9713847054375542
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 3.155406200223499,
54
+ "upper": 3.7974631309509275
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 2.0693869060940213,
58
+ "upper": 2.87911589940389
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.491334999932183,
62
+ "upper": 4.021194886498981
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.6610424452357822,
66
+ "upper": 0.872592170205381
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.6254603839582867,
70
+ "upper": 0.7778024170133803
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.03161394658188025,
74
+ "upper": 0.043074266581485665
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.144492506980896,
84
+ "command_acceleration_rms": 0.12820853292942047,
85
+ "command_jerk_rms": 3.1372432708740234,
86
+ "endpoint_rotation_rmse_deg": 3.8582987785339355,
87
+ "endpoint_translation_rmse_mm": 6.0623393058776855,
88
+ "jerk_rmse": 3.533108949661255,
89
+ "path_rotation_rmse_deg": 2.405137062072754,
90
+ "path_translation_rmse_mm": 3.7594053745269775,
91
+ "residual_rotation_mean_deg": 0.7692350149154663,
92
+ "residual_translation_mean_mm": 0.6971892714500427,
93
+ "velocity_rmse": 0.03634699434041977
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.110406999999999,
98
+ "p95_ms": 8.200868949999998,
99
+ "p99_ms": 10.43797239,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0015241311630234122,
105
+ "endpoint_rotation": 0.10943228751420975,
106
+ "endpoint_translation": 0.002436618087813258,
107
+ "jerk": 0.00226704147644341,
108
+ "path_rotation": 0.061600614339113235,
109
+ "path_translation": 0.0009408381883986294,
110
+ "residual_magnitude": 0.03566799312829971,
111
+ "total": 0.33692434430122375,
112
+ "velocity": 0.04966599866747856
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.1495567262172699,
116
+ "command_acceleration_rms": 0.13441726565361023,
117
+ "command_jerk_rms": 3.3160297870635986,
118
+ "endpoint_rotation_rmse_deg": 3.789569139480591,
119
+ "endpoint_translation_rmse_mm": 6.045598983764648,
120
+ "jerk_rmse": 3.691244125366211,
121
+ "path_rotation_rmse_deg": 2.380246877670288,
122
+ "path_translation_rmse_mm": 3.7566702365875244,
123
+ "residual_rotation_mean_deg": 0.672172486782074,
124
+ "residual_translation_mean_mm": 0.663112223148346,
125
+ "velocity_rmse": 0.03584917262196541
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.15038879215717316,
130
+ "command_acceleration_rms": 0.1392577588558197,
131
+ "command_jerk_rms": 3.4107391834259033,
132
+ "endpoint_rotation_rmse_deg": 4.199632167816162,
133
+ "endpoint_translation_rmse_mm": 6.008756637573242,
134
+ "jerk_rmse": 3.691063404083252,
135
+ "path_rotation_rmse_deg": 2.6454405784606934,
136
+ "path_translation_rmse_mm": 3.743166446685791,
137
+ "residual_rotation_mean_deg": 0.45984548330307007,
138
+ "residual_translation_mean_mm": 0.5546795725822449,
139
+ "velocity_rmse": 0.039565857499837875
140
+ },
141
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
142
+ "acceleration_rmse": 0.15587416291236877,
143
+ "command_acceleration_rms": 0.13095782697200775,
144
+ "command_jerk_rms": 3.4089276790618896,
145
+ "endpoint_rotation_rmse_deg": 3.9784700870513916,
146
+ "endpoint_translation_rmse_mm": 6.946713924407959,
147
+ "jerk_rmse": 3.987175464630127,
148
+ "path_rotation_rmse_deg": 2.506291389465332,
149
+ "path_translation_rmse_mm": 4.335691452026367,
150
+ "residual_rotation_mean_deg": 0.7182945013046265,
151
+ "residual_translation_mean_mm": 0.9254800081253052,
152
+ "velocity_rmse": 0.03756774961948395
153
+ },
154
+ "robotwin_generic::050_bell/base0::candidate_0018": {
155
+ "acceleration_rmse": 0.1444813460111618,
156
+ "command_acceleration_rms": 0.13011473417282104,
157
+ "command_jerk_rms": 3.1705923080444336,
158
+ "endpoint_rotation_rmse_deg": 3.414102554321289,
159
+ "endpoint_translation_rmse_mm": 5.582004547119141,
160
+ "jerk_rmse": 3.5780766010284424,
161
+ "path_rotation_rmse_deg": 2.1180219650268555,
162
+ "path_translation_rmse_mm": 3.4302897453308105,
163
+ "residual_rotation_mean_deg": 0.8319661617279053,
164
+ "residual_translation_mean_mm": 0.7032113075256348,
165
+ "velocity_rmse": 0.03218117728829384
166
+ },
167
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
168
+ "acceleration_rmse": 0.09131758660078049,
169
+ "command_acceleration_rms": 0.09131619334220886,
170
+ "command_jerk_rms": 2.181763172149658,
171
+ "endpoint_rotation_rmse_deg": 6.780344009399414,
172
+ "endpoint_translation_rmse_mm": 7.127496242523193,
173
+ "jerk_rmse": 2.1817626953125,
174
+ "path_rotation_rmse_deg": 4.104794025421143,
175
+ "path_translation_rmse_mm": 4.363109111785889,
176
+ "residual_rotation_mean_deg": 0.5674796104431152,
177
+ "residual_translation_mean_mm": 0.6312280893325806,
178
+ "velocity_rmse": 0.06063513085246086
179
+ },
180
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
181
+ "acceleration_rmse": 0.14704690873622894,
182
+ "command_acceleration_rms": 0.12569846212863922,
183
+ "command_jerk_rms": 3.1258838176727295,
184
+ "endpoint_rotation_rmse_deg": 2.983164072036743,
185
+ "endpoint_translation_rmse_mm": 6.01630973815918,
186
+ "jerk_rmse": 3.631091594696045,
187
+ "path_rotation_rmse_deg": 1.887567400932312,
188
+ "path_translation_rmse_mm": 3.7935662269592285,
189
+ "residual_rotation_mean_deg": 0.9006107449531555,
190
+ "residual_translation_mean_mm": 0.6452526450157166,
191
+ "velocity_rmse": 0.02905953861773014
192
+ },
193
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
194
+ "acceleration_rmse": 0.1694936454296112,
195
+ "command_acceleration_rms": 0.13346831500530243,
196
+ "command_jerk_rms": 3.077533483505249,
197
+ "endpoint_rotation_rmse_deg": 2.9263644218444824,
198
+ "endpoint_translation_rmse_mm": 4.760553359985352,
199
+ "jerk_rmse": 3.9148061275482178,
200
+ "path_rotation_rmse_deg": 1.850063681602478,
201
+ "path_translation_rmse_mm": 2.988938808441162,
202
+ "residual_rotation_mean_deg": 0.9843810796737671,
203
+ "residual_translation_mean_mm": 0.8447241187095642,
204
+ "velocity_rmse": 0.029001181945204735
205
+ },
206
+ "robotwin_generic::071_can/base1::candidate_0084": {
207
+ "acceleration_rmse": 0.13344013690948486,
208
+ "command_acceleration_rms": 0.12478600442409515,
209
+ "command_jerk_rms": 3.220496892929077,
210
+ "endpoint_rotation_rmse_deg": 3.3317511081695557,
211
+ "endpoint_translation_rmse_mm": 6.138230323791504,
212
+ "jerk_rmse": 3.400482416152954,
213
+ "path_rotation_rmse_deg": 2.0716307163238525,
214
+ "path_translation_rmse_mm": 3.78041410446167,
215
+ "residual_rotation_mean_deg": 0.7835702300071716,
216
+ "residual_translation_mean_mm": 0.7476160526275635,
217
+ "velocity_rmse": 0.03152817487716675
218
+ },
219
+ "robotwin_generic::107_soap/base1::candidate_0106": {
220
+ "acceleration_rmse": 0.15871137380599976,
221
+ "command_acceleration_rms": 0.13939915597438812,
222
+ "command_jerk_rms": 3.347158670425415,
223
+ "endpoint_rotation_rmse_deg": 3.444718837738037,
224
+ "endpoint_translation_rmse_mm": 5.598421573638916,
225
+ "jerk_rmse": 3.837700605392456,
226
+ "path_rotation_rmse_deg": 2.184645652770996,
227
+ "path_translation_rmse_mm": 3.4819819927215576,
228
+ "residual_rotation_mean_deg": 0.7184168696403503,
229
+ "residual_translation_mean_mm": 0.6539862155914307,
230
+ "velocity_rmse": 0.032942239195108414
231
+ },
232
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
233
+ "acceleration_rmse": 0.14967872202396393,
234
+ "command_acceleration_rms": 0.13887833058834076,
235
+ "command_jerk_rms": 3.2920944690704346,
236
+ "endpoint_rotation_rmse_deg": 3.6661412715911865,
237
+ "endpoint_translation_rmse_mm": 6.382567405700684,
238
+ "jerk_rmse": 3.575821876525879,
239
+ "path_rotation_rmse_deg": 2.2777774333953857,
240
+ "path_translation_rmse_mm": 3.9174914360046387,
241
+ "residual_rotation_mean_deg": 0.9585511088371277,
242
+ "residual_translation_mean_mm": 0.5685251951217651,
243
+ "velocity_rmse": 0.03464192524552345
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8569679498672484,
250
+ "micro": 0.8517163991928101
251
+ },
252
+ "run_id": "E002-R002",
253
+ "sample_count": 1298,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.15038879215717316,
260
+ "command_acceleration_rms": 0.1392577588558197,
261
+ "command_jerk_rms": 3.4107391834259033,
262
+ "endpoint_rotation_rmse_deg": 4.199632167816162,
263
+ "endpoint_translation_rmse_mm": 6.008756637573242,
264
+ "jerk_rmse": 3.691063404083252,
265
+ "path_rotation_rmse_deg": 2.6454405784606934,
266
+ "path_translation_rmse_mm": 3.743166446685791,
267
+ "residual_rotation_mean_deg": 0.45984548330307007,
268
+ "residual_translation_mean_mm": 0.5546795725822449,
269
+ "velocity_rmse": 0.039565857499837875
270
+ },
271
+ "path_score": 0.9034047603607178,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.14899800717830658,
278
+ "command_acceleration_rms": 0.13108235597610474,
279
+ "command_jerk_rms": 3.251190185546875,
280
+ "endpoint_rotation_rmse_deg": 3.488720178604126,
281
+ "endpoint_translation_rmse_mm": 6.070099353790283,
282
+ "jerk_rmse": 3.6913645267486572,
283
+ "path_rotation_rmse_deg": 2.1851251125335693,
284
+ "path_translation_rmse_mm": 3.765669107437134,
285
+ "residual_rotation_mean_deg": 0.8140876889228821,
286
+ "residual_translation_mean_mm": 0.7355865240097046,
287
+ "velocity_rmse": 0.0331333726644516
288
+ },
289
+ "path_score": 0.8135919332504273,
290
+ "sample_count": 778
291
+ }
292
+ },
293
+ "split": "validation",
294
+ "test_firewall_opened": false
295
+ }
experiments/E002/conditions/c-residual-h16-f186e90df2/run_manifest.yaml ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R002
4
+ experiment_id: E002
5
+ condition_id: c-residual-h16-f186e90df2
6
+ status: completed
7
+ started_at: '2026-08-10T11:31:58.411131+00:00'
8
+ ended_at: '2026-08-10T11:37:02.967653+00:00'
9
+ condition:
10
+ model:
11
+ action_head: residual_chunk
12
+ base_conditioning: E002-R001-profiled-proposal
13
+ horizon_steps: 16
14
+ trajectory_basis: independent_h16
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R002-residual-h16.yaml --data-root
67
+ /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir
68
+ /home/coder/share/experiment-runs/E002/E002-R002 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
69
+ metrics:
70
+ validation:
71
+ loss/total: 0.33688340996189736
72
+ loss/path_translation: 0.000940837494032357
73
+ loss/path_rotation: 0.0615845355097612
74
+ loss/endpoint_translation: 0.0024366182872798806
75
+ loss/endpoint_rotation: 0.10941956785775847
76
+ loss/velocity: 0.04966655885011463
77
+ loss/acceleration: 0.0015241503069759884
78
+ loss/jerk: 0.0022666104369651876
79
+ loss/residual_magnitude: 0.0356679620689071
80
+ metric/path_translation_rmse_mm: 3.753689542573479
81
+ metric/path_rotation_rmse_deg: 2.3661541755100246
82
+ metric/endpoint_translation_rmse_mm: 6.03998319761044
83
+ metric/endpoint_rotation_rmse_deg: 3.7676057393084323
84
+ metric/velocity_rmse: 0.03566566113101315
85
+ metric/acceleration_rmse: 0.1493868921497386
86
+ metric/jerk_rmse: 3.6834481236379943
87
+ metric/command_acceleration_rms: 0.13423365026207293
88
+ metric/command_jerk_rms: 3.309643663500417
89
+ metric/residual_translation_mean_mm: 0.66311021459121
90
+ metric/residual_rotation_mean_deg: 0.6721727473342731
91
+ checkpoint:
92
+ path: /home/coder/share/experiment-runs/E002/E002-R002/best.pt
93
+ sha256: c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9
94
+ selection_metric: loss/total
95
+ selection_split: validation
96
+ artifacts:
97
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R002/metrics.jsonl
98
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R002/evaluation_validation.json
99
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R002/condition.yaml
100
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R002/config_resolved.yaml
101
+ compute:
102
+ wall_seconds: 304.37659596299636
103
+ gpu_hours: 0.08454905443416566
104
+ peak_memory_bytes: 2952011264
experiments/E002/experiment-registry.yaml ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment.v1
2
+ experiment:
3
+ id: E002
4
+ title: se3_bspline_minimum_jerk_recovery_head
5
+ owner: codex
6
+ status: completed
7
+ revision: 2
8
+ experiment_type: confirmatory
9
+ parent_experiment: E001
10
+ revision_history:
11
+ - revision: 1
12
+ frozen_from_git_commit: 71e72ac68e17dfc73e53d5e783efe0c26c7452ed
13
+ reason: User-authorized new structured-head research question; E001 conclusion remains unchanged. Protocol, split,
14
+ conditions, budgets, selection rule, and gates frozen before formal runs.
15
+ - revision: 2
16
+ reason: Results-only conclusion update after the sealed fresh test. Revision 1 condition space, split, selection
17
+ rule, budgets, metrics, and gates are unchanged.
18
+ protocol_revision_1_sha256: fc288032290f2706bd04f58004d5df7166f7823695ad715e44e0bb1bc5bb5417
19
+ gates_changed: false
20
+ research:
21
+ human_prompt_ref: codex-delegation-from-019fdd98-7e76-74f1-bf38-288d75196b2d-2026-08-10
22
+ observation: E001 residual H16 modestly improved path and endpoint error but raised test command jerk RMS from
23
+ 2.311 to 4.300.
24
+ question: Can an explicitly continuous local-tool-frame SE(3) residual curve reduce command jerk versus a same-split
25
+ residual H16 head without materially degrading path or endpoint accuracy?
26
+ hypothesis: Constraining residuals to clamped cubic or true quintic minimum-jerk bases removes independent waypoint
27
+ switching, reducing grouped command jerk by at least 25% versus residual H16 while satisfying pre-registered
28
+ path, endpoint, and latency non-inferiority bounds.
29
+ intervention:
30
+ changed:
31
+ model.action_head:
32
+ - endpoint
33
+ - residual_chunk
34
+ - cubic_bspline
35
+ - minimum_jerk
36
+ model.trajectory_basis:
37
+ - none
38
+ - independent_h16
39
+ - clamped_cubic_bspline_degree3_cp6
40
+ - quintic_minimum_jerk_bezier_cp6
41
+ e001_integrity: E001 remains completed/inconclusive with advance_to_e002=false under its original Pi0.5-conditioned
42
+ Action Chunk definition; this E002 is a separately authorized protocol.
43
+ identifiability_audit:
44
+ duration_seconds: 0.8
45
+ duration_head: excluded because duration is fixed and unidentifiable
46
+ confidence_head: excluded because no confidence supervision or calibration target exists
47
+ implementation: single deterministic residual branch
48
+ baseline:
49
+ condition_id: c-residual-h16-f186e90df2
50
+ preferred_run_id: E002-R002
51
+ comparison: same-split residual H16; endpoint is a second safety reference
52
+ condition_space:
53
+ allowed:
54
+ model.action_head:
55
+ values:
56
+ - endpoint
57
+ - residual_chunk
58
+ - cubic_bspline
59
+ - minimum_jerk
60
+ model.trajectory_basis:
61
+ values:
62
+ - none
63
+ - independent_h16
64
+ - clamped_cubic_bspline_degree3_cp6
65
+ - quintic_minimum_jerk_bezier_cp6
66
+ model.base_conditioning:
67
+ values:
68
+ - none
69
+ - E002-R001-profiled-proposal
70
+ model.horizon_steps:
71
+ values:
72
+ - 1
73
+ - 16
74
+ frozen:
75
+ model.horizon_seconds: 0.8
76
+ model.control_points: 6
77
+ model.backbone.hidden_size: 256
78
+ model.backbone.fusion_layers: 4
79
+ model.decoder_layers: 4
80
+ geometry.residual_curve_frame: initial_local_tool_frame
81
+ geometry.residual_composition: T_cmd[k] = T_base[k] @ Exp_SE3(xi_res(u_k))
82
+ geometry.cubic_basis: open-uniform clamped degree 3; C2 at simple interior knots
83
+ geometry.minimum_jerk_basis: degree-5 Bernstein with controls [0,0,0,q,q,q], phase 10u^3-15u^4+6u^5
84
+ data.hf_repo_id: Shiki42/up-vla-precision-recovery-15k
85
+ data.revision: 77f5f306303361d602ab10759b3a139c67f2cf16
86
+ data.variant_id: validated-physical-recovery-15k-v1
87
+ data.eligible_parent_split: E001 train only
88
+ data.split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
89
+ data.split_namespace: 'E002-source-group-split-v1::'
90
+ training.max_steps: 3000
91
+ training.seed: 42
92
+ evaluation.version: se3-spline-eval-v1
93
+ conditions:
94
+ - run_id: E002-R001
95
+ condition_id: c-endpoint-5d2dff166d
96
+ role: endpoint_baseline
97
+ config: configs/experiments/E002/R001-endpoint.yaml
98
+ status: completed
99
+ - run_id: E002-R002
100
+ condition_id: c-residual-h16-f186e90df2
101
+ role: residual_h16_baseline
102
+ config: configs/experiments/E002/R002-residual-h16.yaml
103
+ status: completed
104
+ - run_id: E002-R003
105
+ condition_id: c-cubic-se3-cp6-43b928e718
106
+ role: candidate
107
+ config: configs/experiments/E002/R003-cubic-se3-cp6.yaml
108
+ status: completed
109
+ - run_id: E002-R004
110
+ condition_id: c-minjerk-se3-cp6-af54a01c9b
111
+ role: candidate
112
+ config: configs/experiments/E002/R004-minjerk-se3-cp6.yaml
113
+ status: completed
114
+ data:
115
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
116
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
117
+ variant_id: validated-physical-recovery-15k-v1
118
+ dataset_hash: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
119
+ split_version: e001-train-only-source-family-stratified-group-hash-e002-v1
120
+ hash_namespace: 'E002-source-group-split-v1::'
121
+ eligible_samples: 11628
122
+ eligible_groups: 89
123
+ split_counts:
124
+ train:
125
+ samples: 8807
126
+ groups: 71
127
+ validation:
128
+ samples: 1298
129
+ groups: 9
130
+ fresh_test:
131
+ samples: 1523
132
+ groups: 9
133
+ holdout_coverage:
134
+ validation:
135
+ - rlbench
136
+ - robotwin_generic
137
+ fresh_test:
138
+ - rlbench
139
+ - robotwin_generic
140
+ exclusion: all 22 E001 validation/test source groups are excluded from E002 confirmatory evidence
141
+ proposal_rule: E002-R001 is trained only on E002 train; frozen proposal generation uses observations/current pose/task
142
+ category only and include_targets=False.
143
+ split_audit:
144
+ path: /home/coder/share/experiment-logs/E002/data-split-audit.json
145
+ sha256: 022dac36e48ce65ea920a789a36e1458b40e78df7f55eabe15504fd8ef0f96be
146
+ target_schema_audit:
147
+ path: /home/coder/share/experiment-logs/E001/action-target-audit.json
148
+ sha256: 16e1b62f9f54fb0ac8a0fab0b7772e66ab6b6bc0d930fc5aa30f53ff19258fa6
149
+ scope: dataset-wide physical/SE(3) target validity only; not reused as E002 model evidence
150
+ evaluation:
151
+ evaluator_version: se3-spline-eval-v1
152
+ primary_metric:
153
+ name: validation_group_macro_path_score
154
+ direction: minimize
155
+ definition: path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5
156
+ core_metric: group_macro_command_jerk_rms
157
+ secondary_metrics:
158
+ - endpoint_translation_rmse_mm
159
+ - endpoint_rotation_rmse_deg
160
+ - velocity_rmse
161
+ - acceleration_rmse
162
+ - jerk_rmse
163
+ - command_acceleration_rms
164
+ - batch1_inference_p95_ms
165
+ group_bootstrap_seed: 20260810
166
+ group_bootstrap_resamples: 10000
167
+ confidence_interval: paired ordinary percentile bootstrap over source groups, 95%
168
+ validation_selection_rule: Select minimum grouped command jerk among candidates satisfying path/endpoint/latency
169
+ guardrails; if none satisfy guardrails, select minimum grouped command jerk among both candidates. Open fresh
170
+ test once for endpoint, residual H16, and that one selected candidate.
171
+ decision_rule:
172
+ advance_requires_on_validation_and_fresh_test:
173
+ - selected_path_score <= 1.02 * residual_h16_path_score
174
+ - selected_command_jerk_rms <= 0.75 * residual_h16_command_jerk_rms
175
+ - selected_command_jerk_rms <= 1.10 * endpoint_command_jerk_rms
176
+ - selected_endpoint_translation_rmse_mm <= endpoint + 1.0
177
+ - selected_endpoint_rotation_rmse_deg <= endpoint + 1.0
178
+ - selected_batch1_gpu_p95_ms < 20.0
179
+ - paired source-group bootstrap 95% CI upper bound for selected/residual command jerk ratio < 1.0
180
+ failure_status: rejected if a safety/non-inferiority gate fails; otherwise inconclusive if evidence is incomplete;
181
+ never relax gates post hoc
182
+ test_firewall: fresh test may be opened exactly once after validation selection, jointly for E002-R001, E002-R002,
183
+ and the selected spline; E001 test is not E002 evidence.
184
+ compute_budget:
185
+ max_scientifically_valid_runs: 4
186
+ max_gpu_hours: 12
187
+ max_training_steps_per_run: 3000
188
+ invalid_implementation_retries: retain and count separately; no fifth scientific condition
189
+ checkpoint_storage:
190
+ provider: huggingface
191
+ repo_id: Shiki42/up-vla-action-chunk-checkpoints
192
+ path_template: experiments/E002/conditions/{condition_id}/
193
+ require_condition_card: true
194
+ upload_policy: upload endpoint baseline, residual H16 baseline, and validation-selected structured condition after
195
+ sealed fresh-test evaluation
196
+ result:
197
+ hypothesis_status: rejected
198
+ advance_to_real_pi05_rollout: false
199
+ selected_condition:
200
+ run_id: E002-R004
201
+ condition_id: c-minjerk-se3-cp6-af54a01c9b
202
+ validation_gate_passed: true
203
+ fresh_test_gate_passed: false
204
+ failed_fresh_test_gates:
205
+ - jerk_reduced_at_least_25pct_vs_residual
206
+ - path_not_worse_than_residual_by_2pct
207
+ summary: Minimum-jerk reduced fresh-test grouped command jerk by 18.50% with paired 95% CI excluding no improvement,
208
+ but missed the 25% effect threshold and worsened residual-H16 path score by 4.18%, above the 2% bound.
209
+ artifacts:
210
+ card: experiments/E002/E002-CARD.md
211
+ summary: experiments/E002/experiment-summary.json
212
+ protocol_freeze: experiments/E002/protocol-freeze-receipt.json
213
+ selection: experiments/E002/selection-validation.json
214
+ firewall: experiments/E002/fresh-test-firewall-receipt.json
experiments/E002/experiment-summary.json ADDED
@@ -0,0 +1,914 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "advance_to_real_pi05_rollout": false,
3
+ "compute": {
4
+ "total_training_gpu_hours": 0.35953789120333063
5
+ },
6
+ "data": {
7
+ "eligible_groups": 89,
8
+ "eligible_samples": 11628,
9
+ "hf_repo_id": "Shiki42/up-vla-precision-recovery-15k",
10
+ "old_e001_holdout_groups_excluded": 22,
11
+ "revision": "77f5f306303361d602ab10759b3a139c67f2cf16",
12
+ "split_audit": {
13
+ "path": "/home/coder/share/experiment-logs/E002/data-split-audit.json",
14
+ "sha256": "022dac36e48ce65ea920a789a36e1458b40e78df7f55eabe15504fd8ef0f96be"
15
+ },
16
+ "split_counts": {
17
+ "fresh_test": {
18
+ "groups": 9,
19
+ "samples": 1523
20
+ },
21
+ "train": {
22
+ "groups": 71,
23
+ "samples": 8807
24
+ },
25
+ "validation": {
26
+ "groups": 9,
27
+ "samples": 1298
28
+ }
29
+ }
30
+ },
31
+ "decision": {
32
+ "next_step": "A new experiment is required for any changed basis, loss, control-point count, or gate; E002 must not be reinterpreted.",
33
+ "real_pi05_rollout_authorized": false,
34
+ "status": "stop"
35
+ },
36
+ "experiment_id": "E002",
37
+ "fresh_test": {
38
+ "firewall_opened_once_for": [
39
+ "E002-R001",
40
+ "E002-R002",
41
+ "E002-R004"
42
+ ],
43
+ "firewall_receipt": {
44
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/firewall.json",
45
+ "sha256": "54a05d0efacd2f3355f3d14d904c275635715b440c9d22ca49a0b88406c48084"
46
+ },
47
+ "gate_passed": false,
48
+ "gates": {
49
+ "batch1_gpu_p95_below_20ms": true,
50
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
51
+ "endpoint_translation_within_endpoint_plus_1mm": true,
52
+ "jerk_not_above_endpoint_by_10pct": true,
53
+ "jerk_reduced_at_least_25pct_vs_residual": false,
54
+ "paired_jerk_ratio_95ci": {
55
+ "lower": 0.7616838102497108,
56
+ "method": "paired-percentile-bootstrap-over-source-groups",
57
+ "point": 0.8149532697598769,
58
+ "resamples": 10000,
59
+ "seed": 20260810,
60
+ "upper": 0.8694310609098554
61
+ },
62
+ "paired_jerk_ratio_95ci_upper_below_1": true,
63
+ "path_not_worse_than_residual_by_2pct": false
64
+ },
65
+ "group_count": 9,
66
+ "metrics": {
67
+ "endpoint": {
68
+ "group_bootstrap_95ci": {
69
+ "confidence": 0.95,
70
+ "group_count": 9,
71
+ "intervals": {
72
+ "acceleration_rmse": {
73
+ "lower": 0.1160274123152097,
74
+ "upper": 0.13443885842959086
75
+ },
76
+ "command_acceleration_rms": {
77
+ "lower": 0.09190107720593611,
78
+ "upper": 0.1158909913773338
79
+ },
80
+ "command_jerk_rms": {
81
+ "lower": 2.229633437262641,
82
+ "upper": 2.815442830324173
83
+ },
84
+ "endpoint_rotation_rmse_deg": {
85
+ "lower": 3.132349097728729,
86
+ "upper": 4.655834012561375
87
+ },
88
+ "endpoint_translation_rmse_mm": {
89
+ "lower": 5.6727268086539375,
90
+ "upper": 6.541422454516093
91
+ },
92
+ "group_macro_path_score": {
93
+ "lower": 0.7736077094740339,
94
+ "upper": 0.9841096719768314
95
+ },
96
+ "jerk_rmse": {
97
+ "lower": 2.810545167658064,
98
+ "upper": 3.28650067448616
99
+ },
100
+ "path_rotation_rmse_deg": {
101
+ "lower": 2.019929062657886,
102
+ "upper": 2.9412611941496527
103
+ },
104
+ "path_translation_rmse_mm": {
105
+ "lower": 3.5075451268090143,
106
+ "upper": 4.077366487847434
107
+ },
108
+ "residual_rotation_mean_deg": {
109
+ "lower": 0.0,
110
+ "upper": 0.0
111
+ },
112
+ "residual_translation_mean_mm": {
113
+ "lower": 0.0,
114
+ "upper": 0.0
115
+ },
116
+ "velocity_rmse": {
117
+ "lower": 0.03070427512543069,
118
+ "upper": 0.043454261496663094
119
+ }
120
+ },
121
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
122
+ "resamples": 10000,
123
+ "seed": 20260810
124
+ },
125
+ "group_macro_path_score": 0.8621170043945312,
126
+ "grouped": {
127
+ "acceleration_rmse": 0.12567059695720673,
128
+ "command_acceleration_rms": 0.10415160655975342,
129
+ "command_jerk_rms": 2.5366859436035156,
130
+ "endpoint_rotation_rmse_deg": 3.7998945713043213,
131
+ "endpoint_translation_rmse_mm": 6.095478534698486,
132
+ "jerk_rmse": 3.062636137008667,
133
+ "path_rotation_rmse_deg": 2.4173941612243652,
134
+ "path_translation_rmse_mm": 3.786381721496582,
135
+ "residual_rotation_mean_deg": 0.0,
136
+ "residual_translation_mean_mm": 0.0,
137
+ "velocity_rmse": 0.03618572652339935
138
+ },
139
+ "latency": {
140
+ "device": "cuda",
141
+ "median_ms": 3.0951405000000003,
142
+ "p95_ms": 3.1367845,
143
+ "p99_ms": 4.168671939999989,
144
+ "repeats": 500,
145
+ "warmup": 50
146
+ }
147
+ },
148
+ "residual_h16": {
149
+ "group_bootstrap_95ci": {
150
+ "confidence": 0.95,
151
+ "group_count": 9,
152
+ "intervals": {
153
+ "acceleration_rmse": {
154
+ "lower": 0.13074185450871786,
155
+ "upper": 0.16456541187233395
156
+ },
157
+ "command_acceleration_rms": {
158
+ "lower": 0.11084854687667554,
159
+ "upper": 0.1475676911158694
160
+ },
161
+ "command_jerk_rms": {
162
+ "lower": 2.763100173738268,
163
+ "upper": 3.5584357129202946
164
+ },
165
+ "endpoint_rotation_rmse_deg": {
166
+ "lower": 2.8186066928837032,
167
+ "upper": 4.241162343819936
168
+ },
169
+ "endpoint_translation_rmse_mm": {
170
+ "lower": 5.567114898893569,
171
+ "upper": 6.269226944446563
172
+ },
173
+ "group_macro_path_score": {
174
+ "lower": 0.7137930199835036,
175
+ "upper": 0.9175362461805343
176
+ },
177
+ "jerk_rmse": {
178
+ "lower": 3.2203344961007434,
179
+ "upper": 3.95932028359837
180
+ },
181
+ "path_rotation_rmse_deg": {
182
+ "lower": 1.7755660480923123,
183
+ "upper": 2.6797972198989655
184
+ },
185
+ "path_translation_rmse_mm": {
186
+ "lower": 3.434561617506875,
187
+ "upper": 3.892301390568415
188
+ },
189
+ "residual_rotation_mean_deg": {
190
+ "lower": 0.6014344857798682,
191
+ "upper": 0.8241439182725218
192
+ },
193
+ "residual_translation_mean_mm": {
194
+ "lower": 0.5710233522786035,
195
+ "upper": 0.6960068250695863
196
+ },
197
+ "velocity_rmse": {
198
+ "lower": 0.027650315987153184,
199
+ "upper": 0.04000451112062567
200
+ }
201
+ },
202
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
203
+ "resamples": 10000,
204
+ "seed": 20260810
205
+ },
206
+ "group_macro_path_score": 0.8018841505050659,
207
+ "grouped": {
208
+ "acceleration_rmse": 0.14657944440841675,
209
+ "command_acceleration_rms": 0.1282518208026886,
210
+ "command_jerk_rms": 3.143152952194214,
211
+ "endpoint_rotation_rmse_deg": 3.4617953300476074,
212
+ "endpoint_translation_rmse_mm": 5.905381202697754,
213
+ "jerk_rmse": 3.5692028999328613,
214
+ "path_rotation_rmse_deg": 2.180569648742676,
215
+ "path_translation_rmse_mm": 3.6577022075653076,
216
+ "residual_rotation_mean_deg": 0.7084853053092957,
217
+ "residual_translation_mean_mm": 0.6295439600944519,
218
+ "velocity_rmse": 0.03316418454051018
219
+ },
220
+ "latency": {
221
+ "device": "cuda",
222
+ "median_ms": 8.124232,
223
+ "p95_ms": 8.236784049999999,
224
+ "p99_ms": 10.43429551,
225
+ "repeats": 500,
226
+ "warmup": 50
227
+ }
228
+ },
229
+ "selected_spline": {
230
+ "group_bootstrap_95ci": {
231
+ "confidence": 0.95,
232
+ "group_count": 9,
233
+ "intervals": {
234
+ "acceleration_rmse": {
235
+ "lower": 0.127791753038764,
236
+ "upper": 0.14594409606522984
237
+ },
238
+ "command_acceleration_rms": {
239
+ "lower": 0.10712421170125405,
240
+ "upper": 0.12841273289587762
241
+ },
242
+ "command_jerk_rms": {
243
+ "lower": 2.2563910881678266,
244
+ "upper": 2.8369754082626764
245
+ },
246
+ "endpoint_rotation_rmse_deg": {
247
+ "lower": 3.017780652973387,
248
+ "upper": 4.63208531803555
249
+ },
250
+ "endpoint_translation_rmse_mm": {
251
+ "lower": 5.5506986127959355,
252
+ "upper": 6.2668231990602274
253
+ },
254
+ "group_macro_path_score": {
255
+ "lower": 0.7442402338319355,
256
+ "upper": 0.9650608261426288
257
+ },
258
+ "jerk_rmse": {
259
+ "lower": 2.8284443749321833,
260
+ "upper": 3.306691540612115
261
+ },
262
+ "path_rotation_rmse_deg": {
263
+ "lower": 1.9337187038527595,
264
+ "upper": 2.9215715991126165
265
+ },
266
+ "path_translation_rmse_mm": {
267
+ "lower": 3.4247371594111127,
268
+ "upper": 3.897781629694833
269
+ },
270
+ "residual_rotation_mean_deg": {
271
+ "lower": 0.49182973437839084,
272
+ "upper": 0.6278564759426647
273
+ },
274
+ "residual_translation_mean_mm": {
275
+ "lower": 0.6452037485109435,
276
+ "upper": 0.8209328550431463
277
+ },
278
+ "velocity_rmse": {
279
+ "lower": 0.030529986860023603,
280
+ "upper": 0.04382059392001894
281
+ }
282
+ },
283
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
284
+ "resamples": 10000,
285
+ "seed": 20260810
286
+ },
287
+ "group_macro_path_score": 0.8353759765625,
288
+ "grouped": {
289
+ "acceleration_rmse": 0.13714249432086945,
290
+ "command_acceleration_rms": 0.11820725351572037,
291
+ "command_jerk_rms": 2.5615224838256836,
292
+ "endpoint_rotation_rmse_deg": 3.70082688331604,
293
+ "endpoint_translation_rmse_mm": 5.904946327209473,
294
+ "jerk_rmse": 3.08172607421875,
295
+ "path_rotation_rmse_deg": 2.3470988273620605,
296
+ "path_translation_rmse_mm": 3.659562110900879,
297
+ "residual_rotation_mean_deg": 0.5654941201210022,
298
+ "residual_translation_mean_mm": 0.723316490650177,
299
+ "velocity_rmse": 0.036055486649274826
300
+ },
301
+ "latency": {
302
+ "device": "cuda",
303
+ "median_ms": 8.102456,
304
+ "p95_ms": 8.21402955,
305
+ "p99_ms": 10.41670332,
306
+ "repeats": 500,
307
+ "warmup": 50
308
+ }
309
+ }
310
+ },
311
+ "sample_count": 1523,
312
+ "summary_receipt": {
313
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/fresh-test-summary.json",
314
+ "sha256": "3751e6ac03749b0bee791baedd4317a269747e35eb92e7b27cd38f2d76373136"
315
+ }
316
+ },
317
+ "hypothesis_status": "rejected",
318
+ "limitations": [
319
+ "Only nine independent groups are present in each E002 holdout.",
320
+ "The selected minimum-jerk head is a single-segment quintic residual curve; it cannot represent multiple interior residual anchors.",
321
+ "Base proposals come from the endpoint recovery model, not real Pi0.5.",
322
+ "Fresh-test group-macro endpoint rotation is influenced by one RLBench group, although grouped bootstrap intervals are reported."
323
+ ],
324
+ "proposal": {
325
+ "endpoint_batch1_gpu_latency": {
326
+ "device": "cuda",
327
+ "median_ms": 3.0951405000000003,
328
+ "p95_ms": 3.1367845,
329
+ "p99_ms": 4.168671939999989,
330
+ "repeats": 500,
331
+ "warmup": 50
332
+ },
333
+ "input_contract": {
334
+ "dataset_include_targets": false,
335
+ "forbidden_fields": [
336
+ "target_pose_world",
337
+ "action_chunk",
338
+ "target_joints",
339
+ "target_gripper"
340
+ ],
341
+ "metadata_fields_for_controller": [
342
+ "current_pose_world",
343
+ "task_category"
344
+ ],
345
+ "observation_modalities": [
346
+ "rgb",
347
+ "depth",
348
+ "point_cloud",
349
+ "proprioception",
350
+ "task_text"
351
+ ]
352
+ },
353
+ "manifest": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
354
+ "sample_count": 11628,
355
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3",
356
+ "target_labels_used": false
357
+ },
358
+ "protocol_freeze": {
359
+ "gates_adjusted_after_results": false,
360
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
361
+ "registry_sha256": "fc288032290f2706bd04f58004d5df7166f7823695ad715e44e0bb1bc5bb5417"
362
+ },
363
+ "reason": "The validation-selected minimum-jerk head failed two frozen fresh-test effect-size/non-inferiority gates: command jerk improved 18.50% rather than >=25%, and path score worsened 4.18% rather than <=2%.",
364
+ "runs": {
365
+ "E002-R001": {
366
+ "checkpoint_sha256": "0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da",
367
+ "code_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
368
+ "command": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py --condition-config configs/experiments/E002/R001-endpoint.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir /home/coder/share/experiment-runs/E002/E002-R001",
369
+ "compute": {
370
+ "gpu_hours": 0.1115510205794433,
371
+ "peak_memory_bytes": 2885576192,
372
+ "wall_seconds": 401.5836740859959
373
+ },
374
+ "condition_id": "c-endpoint-5d2dff166d",
375
+ "environment": {
376
+ "cuda_runtime": "13.0",
377
+ "cudnn_version": 92000,
378
+ "hostname": "lm",
379
+ "os": "Linux-7.0.0-28-generic-x86_64-with-glibc2.35",
380
+ "packages": {
381
+ "Pillow": "11.3.0",
382
+ "PyYAML": "6.0.3",
383
+ "numpy": "1.26.0",
384
+ "torch": "2.13.0"
385
+ },
386
+ "python_executable": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python",
387
+ "python_version": "3.11.15",
388
+ "torch_version": "2.13.0+cu130"
389
+ },
390
+ "hardware": {
391
+ "accelerator_count": 1,
392
+ "accelerators": [
393
+ {
394
+ "compute_capability": "12.0",
395
+ "index": 0,
396
+ "memory_bytes": 101973491712,
397
+ "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
398
+ }
399
+ ],
400
+ "cpu_count": 32
401
+ },
402
+ "manifest": {
403
+ "path": "/home/coder/share/experiment-runs/E002/E002-R001/run_manifest.yaml",
404
+ "sha256": "d273fe41b8dfcbf020c4048aeec3e47296de4bf13bc56bef369df56a8d6c328b"
405
+ },
406
+ "status": "completed"
407
+ },
408
+ "E002-R002": {
409
+ "checkpoint_sha256": "c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9",
410
+ "code_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
411
+ "command": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py --condition-config configs/experiments/E002/R002-residual-h16.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir /home/coder/share/experiment-runs/E002/E002-R002 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals",
412
+ "compute": {
413
+ "gpu_hours": 0.08454905443416566,
414
+ "peak_memory_bytes": 2952011264,
415
+ "wall_seconds": 304.37659596299636
416
+ },
417
+ "condition_id": "c-residual-h16-f186e90df2",
418
+ "environment": {
419
+ "cuda_runtime": "13.0",
420
+ "cudnn_version": 92000,
421
+ "hostname": "lm",
422
+ "os": "Linux-7.0.0-28-generic-x86_64-with-glibc2.35",
423
+ "packages": {
424
+ "Pillow": "11.3.0",
425
+ "PyYAML": "6.0.3",
426
+ "numpy": "1.26.0",
427
+ "torch": "2.13.0"
428
+ },
429
+ "python_executable": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python",
430
+ "python_version": "3.11.15",
431
+ "torch_version": "2.13.0+cu130"
432
+ },
433
+ "hardware": {
434
+ "accelerator_count": 1,
435
+ "accelerators": [
436
+ {
437
+ "compute_capability": "12.0",
438
+ "index": 0,
439
+ "memory_bytes": 101973491712,
440
+ "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
441
+ }
442
+ ],
443
+ "cpu_count": 32
444
+ },
445
+ "manifest": {
446
+ "path": "/home/coder/share/experiment-runs/E002/E002-R002/run_manifest.yaml",
447
+ "sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73"
448
+ },
449
+ "status": "completed"
450
+ },
451
+ "E002-R003": {
452
+ "checkpoint_sha256": "3490f23a0a5bcddb9d1998d08af37318023c8f792b17c68c1562b317dc88554a",
453
+ "code_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
454
+ "command": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py --condition-config configs/experiments/E002/R003-cubic-se3-cp6.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir /home/coder/share/experiment-runs/E002/E002-R003 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals",
455
+ "compute": {
456
+ "gpu_hours": 0.08042972566444304,
457
+ "peak_memory_bytes": 2951959040,
458
+ "wall_seconds": 289.54701239199494
459
+ },
460
+ "condition_id": "c-cubic-se3-cp6-43b928e718",
461
+ "environment": {
462
+ "cuda_runtime": "13.0",
463
+ "cudnn_version": 92000,
464
+ "hostname": "lm",
465
+ "os": "Linux-7.0.0-28-generic-x86_64-with-glibc2.35",
466
+ "packages": {
467
+ "Pillow": "11.3.0",
468
+ "PyYAML": "6.0.3",
469
+ "numpy": "1.26.0",
470
+ "torch": "2.13.0"
471
+ },
472
+ "python_executable": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python",
473
+ "python_version": "3.11.15",
474
+ "torch_version": "2.13.0+cu130"
475
+ },
476
+ "hardware": {
477
+ "accelerator_count": 1,
478
+ "accelerators": [
479
+ {
480
+ "compute_capability": "12.0",
481
+ "index": 0,
482
+ "memory_bytes": 101973491712,
483
+ "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
484
+ }
485
+ ],
486
+ "cpu_count": 32
487
+ },
488
+ "manifest": {
489
+ "path": "/home/coder/share/experiment-runs/E002/E002-R003/run_manifest.yaml",
490
+ "sha256": "6ec30c7b61750729b9750a0fbb86e983c1192defd433e9986c69b6be3fb77298"
491
+ },
492
+ "status": "completed"
493
+ },
494
+ "E002-R004": {
495
+ "checkpoint_sha256": "4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9",
496
+ "code_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
497
+ "command": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py --condition-config configs/experiments/E002/R004-minjerk-se3-cp6.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir /home/coder/share/experiment-runs/E002/E002-R004 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals",
498
+ "compute": {
499
+ "gpu_hours": 0.08300809052527863,
500
+ "peak_memory_bytes": 2951936512,
501
+ "wall_seconds": 298.82912589100306
502
+ },
503
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
504
+ "environment": {
505
+ "cuda_runtime": "13.0",
506
+ "cudnn_version": 92000,
507
+ "hostname": "lm",
508
+ "os": "Linux-7.0.0-28-generic-x86_64-with-glibc2.35",
509
+ "packages": {
510
+ "Pillow": "11.3.0",
511
+ "PyYAML": "6.0.3",
512
+ "numpy": "1.26.0",
513
+ "torch": "2.13.0"
514
+ },
515
+ "python_executable": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python",
516
+ "python_version": "3.11.15",
517
+ "torch_version": "2.13.0+cu130"
518
+ },
519
+ "hardware": {
520
+ "accelerator_count": 1,
521
+ "accelerators": [
522
+ {
523
+ "compute_capability": "12.0",
524
+ "index": 0,
525
+ "memory_bytes": 101973491712,
526
+ "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
527
+ }
528
+ ],
529
+ "cpu_count": 32
530
+ },
531
+ "manifest": {
532
+ "path": "/home/coder/share/experiment-runs/E002/E002-R004/run_manifest.yaml",
533
+ "sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac"
534
+ },
535
+ "status": "completed"
536
+ }
537
+ },
538
+ "schema_version": "upvla.experiment_conclusion.v1",
539
+ "status": "completed",
540
+ "validation": {
541
+ "gate_passed": true,
542
+ "gates": {
543
+ "batch1_gpu_p95_below_20ms": true,
544
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
545
+ "endpoint_translation_within_endpoint_plus_1mm": true,
546
+ "jerk_not_above_endpoint_by_10pct": true,
547
+ "jerk_reduced_at_least_25pct_vs_residual": true,
548
+ "paired_jerk_ratio_95ci": {
549
+ "lower": 0.5945370613267315,
550
+ "method": "paired-percentile-bootstrap-over-source-groups",
551
+ "point": 0.7240625802894608,
552
+ "resamples": 10000,
553
+ "seed": 20260810,
554
+ "upper": 0.8080530623282974
555
+ },
556
+ "paired_jerk_ratio_95ci_upper_below_1": true,
557
+ "path_not_worse_than_residual_by_2pct": true
558
+ },
559
+ "group_count": 9,
560
+ "metrics": {
561
+ "R001": {
562
+ "condition_id": "c-endpoint-5d2dff166d",
563
+ "group_bootstrap_95ci": {
564
+ "confidence": 0.95,
565
+ "group_count": 9,
566
+ "intervals": {
567
+ "acceleration_rmse": {
568
+ "lower": 0.08032658506029595,
569
+ "upper": 0.1295631282652418
570
+ },
571
+ "command_acceleration_rms": {
572
+ "lower": 0.06541380361498643,
573
+ "upper": 0.10663357377052306
574
+ },
575
+ "command_jerk_rms": {
576
+ "lower": 1.6106185335252021,
577
+ "upper": 2.6380579706695344
578
+ },
579
+ "endpoint_rotation_rmse_deg": {
580
+ "lower": 3.3786309235625795,
581
+ "upper": 4.45805601477623
582
+ },
583
+ "endpoint_translation_rmse_mm": {
584
+ "lower": 5.774903562333849,
585
+ "upper": 6.642621228430006
586
+ },
587
+ "group_macro_path_score": {
588
+ "lower": 0.8061851580937703,
589
+ "upper": 0.9451770804325738
590
+ },
591
+ "jerk_rmse": {
592
+ "lower": 1.995983472549253,
593
+ "upper": 3.205968793233236
594
+ },
595
+ "path_rotation_rmse_deg": {
596
+ "lower": 2.1476370354493457,
597
+ "upper": 2.7698281317949296
598
+ },
599
+ "path_translation_rmse_mm": {
600
+ "lower": 3.59573735859659,
601
+ "upper": 4.142488827308019
602
+ },
603
+ "residual_rotation_mean_deg": {
604
+ "lower": 0.0,
605
+ "upper": 0.0
606
+ },
607
+ "residual_translation_mean_mm": {
608
+ "lower": 0.0,
609
+ "upper": 0.0
610
+ },
611
+ "velocity_rmse": {
612
+ "lower": 0.032407565476993715,
613
+ "upper": 0.04119525947090652
614
+ }
615
+ },
616
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
617
+ "resamples": 10000,
618
+ "seed": 20260810
619
+ },
620
+ "group_macro_path_score": 0.8735849380493164,
621
+ "grouped": {
622
+ "acceleration_rmse": 0.11023411154747009,
623
+ "command_acceleration_rms": 0.09015228599309921,
624
+ "command_jerk_rms": 2.220862865447998,
625
+ "endpoint_rotation_rmse_deg": 3.881985664367676,
626
+ "endpoint_translation_rmse_mm": 6.186707973480225,
627
+ "jerk_rmse": 2.729586601257324,
628
+ "path_rotation_rmse_deg": 2.441077709197998,
629
+ "path_translation_rmse_mm": 3.853693962097168,
630
+ "residual_rotation_mean_deg": 0.0,
631
+ "residual_translation_mean_mm": 0.0,
632
+ "velocity_rmse": 0.036509882658720016
633
+ },
634
+ "latency": {
635
+ "device": "cuda",
636
+ "median_ms": 3.0951405000000003,
637
+ "p95_ms": 3.1367845,
638
+ "p99_ms": 4.168671939999989,
639
+ "repeats": 500,
640
+ "warmup": 50
641
+ },
642
+ "receipt": {
643
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R001-validation.json",
644
+ "sha256": "0fda3f48a9f3ac61cdc91d4e13c0cd0d7bbb006c23a4f074d0426313da506b06"
645
+ }
646
+ },
647
+ "R002": {
648
+ "condition_id": "c-residual-h16-f186e90df2",
649
+ "group_bootstrap_95ci": {
650
+ "confidence": 0.95,
651
+ "group_count": 9,
652
+ "intervals": {
653
+ "acceleration_rmse": {
654
+ "lower": 0.12913274992671278,
655
+ "upper": 0.15617000725534227
656
+ },
657
+ "command_acceleration_rms": {
658
+ "lower": 0.11790996123519208,
659
+ "upper": 0.13553185007638402
660
+ },
661
+ "command_jerk_rms": {
662
+ "lower": 2.875136958228217,
663
+ "upper": 3.3160451802942488
664
+ },
665
+ "endpoint_rotation_rmse_deg": {
666
+ "lower": 3.29721736907959,
667
+ "upper": 4.666348218917846
668
+ },
669
+ "endpoint_translation_rmse_mm": {
670
+ "lower": 5.61740255355835,
671
+ "upper": 6.495758643415239
672
+ },
673
+ "group_macro_path_score": {
674
+ "lower": 0.7692517068650987,
675
+ "upper": 0.9713847054375542
676
+ },
677
+ "jerk_rmse": {
678
+ "lower": 3.155406200223499,
679
+ "upper": 3.7974631309509275
680
+ },
681
+ "path_rotation_rmse_deg": {
682
+ "lower": 2.0693869060940213,
683
+ "upper": 2.87911589940389
684
+ },
685
+ "path_translation_rmse_mm": {
686
+ "lower": 3.491334999932183,
687
+ "upper": 4.021194886498981
688
+ },
689
+ "residual_rotation_mean_deg": {
690
+ "lower": 0.6610424452357822,
691
+ "upper": 0.872592170205381
692
+ },
693
+ "residual_translation_mean_mm": {
694
+ "lower": 0.6254603839582867,
695
+ "upper": 0.7778024170133803
696
+ },
697
+ "velocity_rmse": {
698
+ "lower": 0.03161394658188025,
699
+ "upper": 0.043074266581485665
700
+ }
701
+ },
702
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
703
+ "resamples": 10000,
704
+ "seed": 20260810
705
+ },
706
+ "group_macro_path_score": 0.8569679498672484,
707
+ "grouped": {
708
+ "acceleration_rmse": 0.144492506980896,
709
+ "command_acceleration_rms": 0.12820853292942047,
710
+ "command_jerk_rms": 3.1372432708740234,
711
+ "endpoint_rotation_rmse_deg": 3.8582987785339355,
712
+ "endpoint_translation_rmse_mm": 6.0623393058776855,
713
+ "jerk_rmse": 3.533108949661255,
714
+ "path_rotation_rmse_deg": 2.405137062072754,
715
+ "path_translation_rmse_mm": 3.7594053745269775,
716
+ "residual_rotation_mean_deg": 0.7692350149154663,
717
+ "residual_translation_mean_mm": 0.6971892714500427,
718
+ "velocity_rmse": 0.03634699434041977
719
+ },
720
+ "latency": {
721
+ "device": "cuda",
722
+ "median_ms": 8.110406999999999,
723
+ "p95_ms": 8.200868949999998,
724
+ "p99_ms": 10.43797239,
725
+ "repeats": 500,
726
+ "warmup": 50
727
+ },
728
+ "receipt": {
729
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R002-validation.json",
730
+ "sha256": "17a116944f1967d60d2427975690f635f4886180578b92b969be7d9d756c84a1"
731
+ }
732
+ },
733
+ "R003": {
734
+ "condition_id": "c-cubic-se3-cp6-43b928e718",
735
+ "group_bootstrap_95ci": {
736
+ "confidence": 0.95,
737
+ "group_count": 9,
738
+ "intervals": {
739
+ "acceleration_rmse": {
740
+ "lower": 0.11298756880892648,
741
+ "upper": 0.1394385862888561
742
+ },
743
+ "command_acceleration_rms": {
744
+ "lower": 0.09995081027348836,
745
+ "upper": 0.11602214109152555
746
+ },
747
+ "command_jerk_rms": {
748
+ "lower": 1.8295588115851085,
749
+ "upper": 2.6620728174845376
750
+ },
751
+ "endpoint_rotation_rmse_deg": {
752
+ "lower": 3.2371414442857107,
753
+ "upper": 4.614357245630688
754
+ },
755
+ "endpoint_translation_rmse_mm": {
756
+ "lower": 5.5499372945891485,
757
+ "upper": 6.477358905474345
758
+ },
759
+ "group_macro_path_score": {
760
+ "lower": 0.7578355182541741,
761
+ "upper": 0.9678520631127886
762
+ },
763
+ "jerk_rmse": {
764
+ "lower": 2.2059771882163153,
765
+ "upper": 3.228559059566922
766
+ },
767
+ "path_rotation_rmse_deg": {
768
+ "lower": 2.0445298645231462,
769
+ "upper": 2.8757890260881847
770
+ },
771
+ "path_translation_rmse_mm": {
772
+ "lower": 3.434231901168823,
773
+ "upper": 3.9773682713508607
774
+ },
775
+ "residual_rotation_mean_deg": {
776
+ "lower": 0.7555553118387858,
777
+ "upper": 0.9298186732663049
778
+ },
779
+ "residual_translation_mean_mm": {
780
+ "lower": 0.8847316967116462,
781
+ "upper": 1.1563906035489506
782
+ },
783
+ "velocity_rmse": {
784
+ "lower": 0.031309114784623185,
785
+ "upper": 0.042778491559955806
786
+ }
787
+ },
788
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
789
+ "resamples": 10000,
790
+ "seed": 20260810
791
+ },
792
+ "group_macro_path_score": 0.8477689743041992,
793
+ "grouped": {
794
+ "acceleration_rmse": 0.1268954873085022,
795
+ "command_acceleration_rms": 0.10837496817111969,
796
+ "command_jerk_rms": 2.313187599182129,
797
+ "endpoint_rotation_rmse_deg": 3.8112881183624268,
798
+ "endpoint_translation_rmse_mm": 6.002523422241211,
799
+ "jerk_rmse": 2.8165056705474854,
800
+ "path_rotation_rmse_deg": 2.38803768157959,
801
+ "path_translation_rmse_mm": 3.7016143798828125,
802
+ "residual_rotation_mean_deg": 0.84266597032547,
803
+ "residual_translation_mean_mm": 1.0173211097717285,
804
+ "velocity_rmse": 0.03602871671319008
805
+ },
806
+ "latency": {
807
+ "device": "cuda",
808
+ "median_ms": 8.1211475,
809
+ "p95_ms": 8.21552085,
810
+ "p99_ms": 8.44476369,
811
+ "repeats": 500,
812
+ "warmup": 50
813
+ },
814
+ "receipt": {
815
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R003-validation.json",
816
+ "sha256": "237728c0a5594821ef420a5ec4452943767845a7058ccf32f7dc6515ab7cd330"
817
+ }
818
+ },
819
+ "R004": {
820
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
821
+ "group_bootstrap_95ci": {
822
+ "confidence": 0.95,
823
+ "group_count": 9,
824
+ "intervals": {
825
+ "acceleration_rmse": {
826
+ "lower": 0.10308757486442725,
827
+ "upper": 0.1397645959837569
828
+ },
829
+ "command_acceleration_rms": {
830
+ "lower": 0.08936153536455499,
831
+ "upper": 0.11839342810627487
832
+ },
833
+ "command_jerk_rms": {
834
+ "lower": 1.719706772930092,
835
+ "upper": 2.6594989677270253
836
+ },
837
+ "endpoint_rotation_rmse_deg": {
838
+ "lower": 3.205244329240587,
839
+ "upper": 4.427248266008165
840
+ },
841
+ "endpoint_translation_rmse_mm": {
842
+ "lower": 5.687558265527089,
843
+ "upper": 6.8257540742556255
844
+ },
845
+ "group_macro_path_score": {
846
+ "lower": 0.7683746231926811,
847
+ "upper": 0.9666738802856869
848
+ },
849
+ "jerk_rmse": {
850
+ "lower": 2.1001228921943245,
851
+ "upper": 3.2242833753426865
852
+ },
853
+ "path_rotation_rmse_deg": {
854
+ "lower": 2.028298435277409,
855
+ "upper": 2.7444795823759502
856
+ },
857
+ "path_translation_rmse_mm": {
858
+ "lower": 3.546643653843138,
859
+ "upper": 4.226622504658169
860
+ },
861
+ "residual_rotation_mean_deg": {
862
+ "lower": 0.5349525012903743,
863
+ "upper": 0.5974235445261001
864
+ },
865
+ "residual_translation_mean_mm": {
866
+ "lower": 0.7040380966332224,
867
+ "upper": 0.9369111968411339
868
+ },
869
+ "velocity_rmse": {
870
+ "lower": 0.03159253713157442,
871
+ "upper": 0.041467026931544144
872
+ }
873
+ },
874
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
875
+ "resamples": 10000,
876
+ "seed": 20260810
877
+ },
878
+ "group_macro_path_score": 0.8557925224304199,
879
+ "grouped": {
880
+ "acceleration_rmse": 0.12487276643514633,
881
+ "command_acceleration_rms": 0.10622868686914444,
882
+ "command_jerk_rms": 2.2715606689453125,
883
+ "endpoint_rotation_rmse_deg": 3.7452497482299805,
884
+ "endpoint_translation_rmse_mm": 6.209352493286133,
885
+ "jerk_rmse": 2.7772092819213867,
886
+ "path_rotation_rmse_deg": 2.348088264465332,
887
+ "path_translation_rmse_mm": 3.861748695373535,
888
+ "residual_rotation_mean_deg": 0.5649371147155762,
889
+ "residual_translation_mean_mm": 0.8120338320732117,
890
+ "velocity_rmse": 0.0359807126224041
891
+ },
892
+ "latency": {
893
+ "device": "cuda",
894
+ "median_ms": 8.1037045,
895
+ "p95_ms": 8.980220799999982,
896
+ "p99_ms": 10.40966553,
897
+ "repeats": 500,
898
+ "warmup": 50
899
+ },
900
+ "receipt": {
901
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R004-validation.json",
902
+ "sha256": "7c451bf90cb7e2f1ea50a3dff1647bca8a9cea6d4adf9003733d8140e6beb741"
903
+ }
904
+ }
905
+ },
906
+ "sample_count": 1298,
907
+ "selected_condition_id": "c-minjerk-se3-cp6-af54a01c9b",
908
+ "selected_run_id": "E002-R004",
909
+ "selection_receipt": {
910
+ "path": "/home/coder/share/experiment-artifacts/E002/selection-validation.json",
911
+ "sha256": "8c3e64ef2f874fda8ec4d263d389dacb5abb1f5b2253bd4d47b9cce9fd7a7ec4"
912
+ }
913
+ }
914
+ }
experiments/E002/fresh-test-firewall-receipt.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "advance_to_real_pi05_rollout": false,
3
+ "allowed_runs": [
4
+ "E002-R001",
5
+ "E002-R002",
6
+ "E002-R004"
7
+ ],
8
+ "commands": {
9
+ "endpoint": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python /home/coder/share/UP-VLA-se3-bspline-E002/scripts/evaluate_action_chunk_recovery.py --condition-config /home/coder/share/UP-VLA-se3-bspline-E002/configs/experiments/E002/R001-endpoint.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --split test --prediction-source base_proposal --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals --output /home/coder/share/experiment-artifacts/E002/fresh-test-sealed.opening/endpoint-test.json --open-test-firewall",
10
+ "residual_h16": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python /home/coder/share/UP-VLA-se3-bspline-E002/scripts/evaluate_action_chunk_recovery.py --condition-config /home/coder/share/UP-VLA-se3-bspline-E002/configs/experiments/E002/R002-residual-h16.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --split test --prediction-source checkpoint --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals --output /home/coder/share/experiment-artifacts/E002/fresh-test-sealed.opening/residual_h16-test.json --open-test-firewall --checkpoint /home/coder/share/experiment-runs/E002/E002-R002/best.pt",
11
+ "selected_spline": "/home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python /home/coder/share/UP-VLA-se3-bspline-E002/scripts/evaluate_action_chunk_recovery.py --condition-config /home/coder/share/UP-VLA-se3-bspline-E002/configs/experiments/E002/R004-minjerk-se3-cp6.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --split test --prediction-source checkpoint --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals --output /home/coder/share/experiment-artifacts/E002/fresh-test-sealed.opening/selected_spline-test.json --open-test-firewall --checkpoint /home/coder/share/experiment-runs/E002/E002-R004/best.pt"
12
+ },
13
+ "e001_test_reused": false,
14
+ "experiment": {
15
+ "path": "/home/coder/share/UP-VLA-se3-bspline-E002/experiments/E002-se3-bspline-minjerk-recovery.yaml",
16
+ "sha256": "fc288032290f2706bd04f58004d5df7166f7823695ad715e44e0bb1bc5bb5417"
17
+ },
18
+ "fresh_split": "E002 test groups from the frozen E001-train-only re-split",
19
+ "hypothesis_status": "rejected",
20
+ "schema_version": "precision.e002.test_firewall.v1",
21
+ "selection": {
22
+ "path": "/home/coder/share/experiment-artifacts/E002/selection-validation.json",
23
+ "sha256": "8c3e64ef2f874fda8ec4d263d389dacb5abb1f5b2253bd4d47b9cce9fd7a7ec4"
24
+ },
25
+ "status": "sealed",
26
+ "summary": {
27
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/fresh-test-summary.json",
28
+ "sha256": "3751e6ac03749b0bee791baedd4317a269747e35eb92e7b27cd38f2d76373136"
29
+ }
30
+ }
experiments/E002/fresh-test-summary.json ADDED
@@ -0,0 +1,290 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "advance_to_real_pi05_rollout": false,
3
+ "experiment_id": "E002",
4
+ "firewall_opened_once_for": [
5
+ "E002-R001",
6
+ "E002-R002",
7
+ "E002-R004"
8
+ ],
9
+ "fresh_test_gate_passed": false,
10
+ "fresh_test_gates": {
11
+ "batch1_gpu_p95_below_20ms": true,
12
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
13
+ "endpoint_translation_within_endpoint_plus_1mm": true,
14
+ "jerk_not_above_endpoint_by_10pct": true,
15
+ "jerk_reduced_at_least_25pct_vs_residual": false,
16
+ "paired_jerk_ratio_95ci": {
17
+ "lower": 0.7616838102497108,
18
+ "method": "paired-percentile-bootstrap-over-source-groups",
19
+ "point": 0.8149532697598769,
20
+ "resamples": 10000,
21
+ "seed": 20260810,
22
+ "upper": 0.8694310609098554
23
+ },
24
+ "paired_jerk_ratio_95ci_upper_below_1": true,
25
+ "path_not_worse_than_residual_by_2pct": false
26
+ },
27
+ "hypothesis_status": "rejected",
28
+ "metrics": {
29
+ "endpoint": {
30
+ "group_bootstrap_95ci": {
31
+ "confidence": 0.95,
32
+ "group_count": 9,
33
+ "intervals": {
34
+ "acceleration_rmse": {
35
+ "lower": 0.1160274123152097,
36
+ "upper": 0.13443885842959086
37
+ },
38
+ "command_acceleration_rms": {
39
+ "lower": 0.09190107720593611,
40
+ "upper": 0.1158909913773338
41
+ },
42
+ "command_jerk_rms": {
43
+ "lower": 2.229633437262641,
44
+ "upper": 2.815442830324173
45
+ },
46
+ "endpoint_rotation_rmse_deg": {
47
+ "lower": 3.132349097728729,
48
+ "upper": 4.655834012561375
49
+ },
50
+ "endpoint_translation_rmse_mm": {
51
+ "lower": 5.6727268086539375,
52
+ "upper": 6.541422454516093
53
+ },
54
+ "group_macro_path_score": {
55
+ "lower": 0.7736077094740339,
56
+ "upper": 0.9841096719768314
57
+ },
58
+ "jerk_rmse": {
59
+ "lower": 2.810545167658064,
60
+ "upper": 3.28650067448616
61
+ },
62
+ "path_rotation_rmse_deg": {
63
+ "lower": 2.019929062657886,
64
+ "upper": 2.9412611941496527
65
+ },
66
+ "path_translation_rmse_mm": {
67
+ "lower": 3.5075451268090143,
68
+ "upper": 4.077366487847434
69
+ },
70
+ "residual_rotation_mean_deg": {
71
+ "lower": 0.0,
72
+ "upper": 0.0
73
+ },
74
+ "residual_translation_mean_mm": {
75
+ "lower": 0.0,
76
+ "upper": 0.0
77
+ },
78
+ "velocity_rmse": {
79
+ "lower": 0.03070427512543069,
80
+ "upper": 0.043454261496663094
81
+ }
82
+ },
83
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
84
+ "resamples": 10000,
85
+ "seed": 20260810
86
+ },
87
+ "group_macro_path_score": 0.8621170043945312,
88
+ "grouped": {
89
+ "acceleration_rmse": 0.12567059695720673,
90
+ "command_acceleration_rms": 0.10415160655975342,
91
+ "command_jerk_rms": 2.5366859436035156,
92
+ "endpoint_rotation_rmse_deg": 3.7998945713043213,
93
+ "endpoint_translation_rmse_mm": 6.095478534698486,
94
+ "jerk_rmse": 3.062636137008667,
95
+ "path_rotation_rmse_deg": 2.4173941612243652,
96
+ "path_translation_rmse_mm": 3.786381721496582,
97
+ "residual_rotation_mean_deg": 0.0,
98
+ "residual_translation_mean_mm": 0.0,
99
+ "velocity_rmse": 0.03618572652339935
100
+ },
101
+ "latency": {
102
+ "device": "cuda",
103
+ "median_ms": 3.0951405000000003,
104
+ "p95_ms": 3.1367845,
105
+ "p99_ms": 4.168671939999989,
106
+ "repeats": 500,
107
+ "warmup": 50
108
+ }
109
+ },
110
+ "residual_h16": {
111
+ "group_bootstrap_95ci": {
112
+ "confidence": 0.95,
113
+ "group_count": 9,
114
+ "intervals": {
115
+ "acceleration_rmse": {
116
+ "lower": 0.13074185450871786,
117
+ "upper": 0.16456541187233395
118
+ },
119
+ "command_acceleration_rms": {
120
+ "lower": 0.11084854687667554,
121
+ "upper": 0.1475676911158694
122
+ },
123
+ "command_jerk_rms": {
124
+ "lower": 2.763100173738268,
125
+ "upper": 3.5584357129202946
126
+ },
127
+ "endpoint_rotation_rmse_deg": {
128
+ "lower": 2.8186066928837032,
129
+ "upper": 4.241162343819936
130
+ },
131
+ "endpoint_translation_rmse_mm": {
132
+ "lower": 5.567114898893569,
133
+ "upper": 6.269226944446563
134
+ },
135
+ "group_macro_path_score": {
136
+ "lower": 0.7137930199835036,
137
+ "upper": 0.9175362461805343
138
+ },
139
+ "jerk_rmse": {
140
+ "lower": 3.2203344961007434,
141
+ "upper": 3.95932028359837
142
+ },
143
+ "path_rotation_rmse_deg": {
144
+ "lower": 1.7755660480923123,
145
+ "upper": 2.6797972198989655
146
+ },
147
+ "path_translation_rmse_mm": {
148
+ "lower": 3.434561617506875,
149
+ "upper": 3.892301390568415
150
+ },
151
+ "residual_rotation_mean_deg": {
152
+ "lower": 0.6014344857798682,
153
+ "upper": 0.8241439182725218
154
+ },
155
+ "residual_translation_mean_mm": {
156
+ "lower": 0.5710233522786035,
157
+ "upper": 0.6960068250695863
158
+ },
159
+ "velocity_rmse": {
160
+ "lower": 0.027650315987153184,
161
+ "upper": 0.04000451112062567
162
+ }
163
+ },
164
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
165
+ "resamples": 10000,
166
+ "seed": 20260810
167
+ },
168
+ "group_macro_path_score": 0.8018841505050659,
169
+ "grouped": {
170
+ "acceleration_rmse": 0.14657944440841675,
171
+ "command_acceleration_rms": 0.1282518208026886,
172
+ "command_jerk_rms": 3.143152952194214,
173
+ "endpoint_rotation_rmse_deg": 3.4617953300476074,
174
+ "endpoint_translation_rmse_mm": 5.905381202697754,
175
+ "jerk_rmse": 3.5692028999328613,
176
+ "path_rotation_rmse_deg": 2.180569648742676,
177
+ "path_translation_rmse_mm": 3.6577022075653076,
178
+ "residual_rotation_mean_deg": 0.7084853053092957,
179
+ "residual_translation_mean_mm": 0.6295439600944519,
180
+ "velocity_rmse": 0.03316418454051018
181
+ },
182
+ "latency": {
183
+ "device": "cuda",
184
+ "median_ms": 8.124232,
185
+ "p95_ms": 8.236784049999999,
186
+ "p99_ms": 10.43429551,
187
+ "repeats": 500,
188
+ "warmup": 50
189
+ }
190
+ },
191
+ "selected_spline": {
192
+ "group_bootstrap_95ci": {
193
+ "confidence": 0.95,
194
+ "group_count": 9,
195
+ "intervals": {
196
+ "acceleration_rmse": {
197
+ "lower": 0.127791753038764,
198
+ "upper": 0.14594409606522984
199
+ },
200
+ "command_acceleration_rms": {
201
+ "lower": 0.10712421170125405,
202
+ "upper": 0.12841273289587762
203
+ },
204
+ "command_jerk_rms": {
205
+ "lower": 2.2563910881678266,
206
+ "upper": 2.8369754082626764
207
+ },
208
+ "endpoint_rotation_rmse_deg": {
209
+ "lower": 3.017780652973387,
210
+ "upper": 4.63208531803555
211
+ },
212
+ "endpoint_translation_rmse_mm": {
213
+ "lower": 5.5506986127959355,
214
+ "upper": 6.2668231990602274
215
+ },
216
+ "group_macro_path_score": {
217
+ "lower": 0.7442402338319355,
218
+ "upper": 0.9650608261426288
219
+ },
220
+ "jerk_rmse": {
221
+ "lower": 2.8284443749321833,
222
+ "upper": 3.306691540612115
223
+ },
224
+ "path_rotation_rmse_deg": {
225
+ "lower": 1.9337187038527595,
226
+ "upper": 2.9215715991126165
227
+ },
228
+ "path_translation_rmse_mm": {
229
+ "lower": 3.4247371594111127,
230
+ "upper": 3.897781629694833
231
+ },
232
+ "residual_rotation_mean_deg": {
233
+ "lower": 0.49182973437839084,
234
+ "upper": 0.6278564759426647
235
+ },
236
+ "residual_translation_mean_mm": {
237
+ "lower": 0.6452037485109435,
238
+ "upper": 0.8209328550431463
239
+ },
240
+ "velocity_rmse": {
241
+ "lower": 0.030529986860023603,
242
+ "upper": 0.04382059392001894
243
+ }
244
+ },
245
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
246
+ "resamples": 10000,
247
+ "seed": 20260810
248
+ },
249
+ "group_macro_path_score": 0.8353759765625,
250
+ "grouped": {
251
+ "acceleration_rmse": 0.13714249432086945,
252
+ "command_acceleration_rms": 0.11820725351572037,
253
+ "command_jerk_rms": 2.5615224838256836,
254
+ "endpoint_rotation_rmse_deg": 3.70082688331604,
255
+ "endpoint_translation_rmse_mm": 5.904946327209473,
256
+ "jerk_rmse": 3.08172607421875,
257
+ "path_rotation_rmse_deg": 2.3470988273620605,
258
+ "path_translation_rmse_mm": 3.659562110900879,
259
+ "residual_rotation_mean_deg": 0.5654941201210022,
260
+ "residual_translation_mean_mm": 0.723316490650177,
261
+ "velocity_rmse": 0.036055486649274826
262
+ },
263
+ "latency": {
264
+ "device": "cuda",
265
+ "median_ms": 8.102456,
266
+ "p95_ms": 8.21402955,
267
+ "p99_ms": 10.41670332,
268
+ "repeats": 500,
269
+ "warmup": 50
270
+ }
271
+ }
272
+ },
273
+ "reports": {
274
+ "endpoint": {
275
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/endpoint-test.json",
276
+ "sha256": "6bac903b2a7a547f753130a77604bafda1e4f76a8138396734c0bd98bd7187df"
277
+ },
278
+ "residual_h16": {
279
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/residual_h16-test.json",
280
+ "sha256": "390e69b77677e4e6096afa2eefde8edb633f88558746f715ecbfb9ed03102742"
281
+ },
282
+ "selected_spline": {
283
+ "path": "/home/coder/share/experiment-artifacts/E002/fresh-test-sealed/selected_spline-test.json",
284
+ "sha256": "8c32ee98a9039c7a5f426f44c5aec1eec00d65e06347a15d2ed9899387d2e4c7"
285
+ }
286
+ },
287
+ "schema_version": "precision.e002.fresh_test_summary.v1",
288
+ "selected_condition_id": "c-minjerk-se3-cp6-af54a01c9b",
289
+ "validation_gate_passed": true
290
+ }
experiments/E002/protocol-freeze-receipt.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "condition_ids": [
3
+ "c-endpoint-5d2dff166d",
4
+ "c-residual-h16-f186e90df2",
5
+ "c-cubic-se3-cp6-43b928e718",
6
+ "c-minjerk-se3-cp6-af54a01c9b"
7
+ ],
8
+ "experiment_id": "E002",
9
+ "formal_runs": [
10
+ "E002-R001",
11
+ "E002-R002",
12
+ "E002-R003",
13
+ "E002-R004"
14
+ ],
15
+ "gates_adjusted_after_results": false,
16
+ "git_branch": "shuyuan/se3-bspline-recovery-e002",
17
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
18
+ "registry_path": "experiments/E002-se3-bspline-minjerk-recovery.yaml",
19
+ "registry_sha256": "fc288032290f2706bd04f58004d5df7166f7823695ad715e44e0bb1bc5bb5417",
20
+ "schema_version": "precision.e002.protocol_freeze.v1",
21
+ "split_audit_sha256": "022dac36e48ce65ea920a789a36e1458b40e78df7f55eabe15504fd8ef0f96be"
22
+ }
experiments/E002/runs/E002-R001/condition_config.yaml ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.condition_config.v1
2
+ experiment_id: E002
3
+ run_id: E002-R001
4
+ condition_id: c-endpoint-5d2dff166d
5
+ condition_summary: endpoint
6
+ condition:
7
+ model:
8
+ action_head: endpoint
9
+ base_conditioning: none
10
+ horizon_steps: 1
11
+ trajectory_basis: none
12
+ resolved_config:
13
+ data:
14
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
15
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
16
+ variant_id: validated-physical-recovery-15k-v1
17
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
18
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
19
+ sample_count: 15000
20
+ eligible_sample_count: 11628
21
+ eligible_group_count: 89
22
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
23
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
24
+ split_namespace: 'E002-source-group-split-v1::'
25
+ validation_fraction: 0.1
26
+ test_fraction: 0.1
27
+ geometry:
28
+ cumulative_waypoint_representation: current_tool_relative_pose
29
+ step_representation: integrated_body_twist
30
+ step_integration: right_se3_exponential
31
+ residual_curve_frame: initial_local_tool_frame
32
+ residual_composition: base_right_multiply_se3_exponential
33
+ structured_translation: se3_tangent_translation_control_points_m
34
+ structured_rotation: so3_lie_algebra_control_points_rad
35
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
36
+ residual_boundary_at_t0: identity
37
+ model:
38
+ backbone:
39
+ hidden_size: 256
40
+ fusion_layers: 4
41
+ fusion_heads: 8
42
+ fusion_mlp_ratio: 4
43
+ dropout: 0.1
44
+ vision_backend: conv
45
+ text_backend: embedding
46
+ pretrained_model_name: google/siglip-base-patch16-224
47
+ freeze_pretrained: true
48
+ image_size: 224
49
+ vision_tokens: 64
50
+ depth_tokens: 36
51
+ point_tokens: 32
52
+ point_samples: 256
53
+ max_text_tokens: 32
54
+ text_vocab_size: 32768
55
+ max_joints: 7
56
+ aloha_joints: 6
57
+ panda_joints: 7
58
+ translation_scale_m: 0.05
59
+ rotation_scale_rad: 0.5235987755982988
60
+ joint_scale_rad: 0.5
61
+ horizon_steps: 16
62
+ horizon_seconds: 0.8
63
+ decoder_layers: 4
64
+ decoder_heads: 8
65
+ decoder_mlp_ratio: 4
66
+ control_points: 6
67
+ maximum_step_translation_m: 0.01
68
+ maximum_step_rotation_rad: 0.2
69
+ maximum_residual_translation_m: 0.05
70
+ maximum_residual_rotation_rad: 0.52
71
+ action_head: endpoint
72
+ base_conditioning: none
73
+ trajectory_basis: none
74
+ loss:
75
+ translation: 1.0
76
+ rotation_vector: 1.0
77
+ rotation_geodesic: 0.25
78
+ joint: 0.5
79
+ gripper: 0.1
80
+ training:
81
+ seed: 42
82
+ batch_size: 64
83
+ gradient_accumulation_steps: 1
84
+ epochs: 100
85
+ max_steps: 3000
86
+ learning_rate: 0.0003
87
+ weight_decay: 0.05
88
+ warmup_fraction: 0.05
89
+ gradient_clip_norm: 1.0
90
+ precision: bf16
91
+ workers: 8
92
+ log_every_steps: 10
93
+ validation_every_steps: 250
94
+ checkpoint_every_steps: 500
95
+ evaluation:
96
+ version: se3-spline-eval-v1
97
+ trajectory_loss:
98
+ path_translation: 1.0
99
+ path_rotation: 1.0
100
+ endpoint_translation: 2.0
101
+ endpoint_rotation: 2.0
102
+ velocity: 1.0
103
+ acceleration: 0.1
104
+ jerk: 0.05
105
+ residual_magnitude: 0.02
106
+ selection_split: validation
107
+ test_firewall_opened: false
108
+ group_bootstrap_seed: 20260810
109
+ group_bootstrap_resamples: 10000
experiments/E002/runs/E002-R001/evaluation_validation.json ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": null,
7
+ "checkpoint_run": null,
8
+ "condition_id": "c-endpoint-5d2dff166d",
9
+ "data_extraction_receipt": {
10
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
11
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
12
+ },
13
+ "experiment_id": "E002",
14
+ "group_bootstrap_95ci": {
15
+ "confidence": 0.95,
16
+ "group_count": 9,
17
+ "intervals": {
18
+ "acceleration_rmse": {
19
+ "lower": 0.08032658506029595,
20
+ "upper": 0.1295631282652418
21
+ },
22
+ "command_acceleration_rms": {
23
+ "lower": 0.06541380361498643,
24
+ "upper": 0.10663357377052306
25
+ },
26
+ "command_jerk_rms": {
27
+ "lower": 1.6106185335252021,
28
+ "upper": 2.6380579706695344
29
+ },
30
+ "endpoint_rotation_rmse_deg": {
31
+ "lower": 3.3786309235625795,
32
+ "upper": 4.45805601477623
33
+ },
34
+ "endpoint_translation_rmse_mm": {
35
+ "lower": 5.774903562333849,
36
+ "upper": 6.642621228430006
37
+ },
38
+ "group_macro_path_score": {
39
+ "lower": 0.8061851580937703,
40
+ "upper": 0.9451770804325738
41
+ },
42
+ "jerk_rmse": {
43
+ "lower": 1.995983472549253,
44
+ "upper": 3.205968793233236
45
+ },
46
+ "path_rotation_rmse_deg": {
47
+ "lower": 2.1476370354493457,
48
+ "upper": 2.7698281317949296
49
+ },
50
+ "path_translation_rmse_mm": {
51
+ "lower": 3.59573735859659,
52
+ "upper": 4.142488827308019
53
+ },
54
+ "residual_rotation_mean_deg": {
55
+ "lower": 0.0,
56
+ "upper": 0.0
57
+ },
58
+ "residual_translation_mean_mm": {
59
+ "lower": 0.0,
60
+ "upper": 0.0
61
+ },
62
+ "velocity_rmse": {
63
+ "lower": 0.032407565476993715,
64
+ "upper": 0.04119525947090652
65
+ }
66
+ },
67
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
68
+ "resamples": 10000,
69
+ "seed": 20260810
70
+ },
71
+ "group_count": 9,
72
+ "grouped_metrics": {
73
+ "acceleration_rmse": 0.11023411154747009,
74
+ "command_acceleration_rms": 0.09015228599309921,
75
+ "command_jerk_rms": 2.220862865447998,
76
+ "endpoint_rotation_rmse_deg": 3.881985664367676,
77
+ "endpoint_translation_rmse_mm": 6.186707973480225,
78
+ "jerk_rmse": 2.729586601257324,
79
+ "path_rotation_rmse_deg": 2.441077709197998,
80
+ "path_translation_rmse_mm": 3.853693962097168,
81
+ "residual_rotation_mean_deg": 0.0,
82
+ "residual_translation_mean_mm": 0.0,
83
+ "velocity_rmse": 0.036509882658720016
84
+ },
85
+ "latency": null,
86
+ "losses": {
87
+ "acceleration": 0.001078446046449244,
88
+ "endpoint_rotation": 0.11553905159235,
89
+ "endpoint_translation": 0.002643197774887085,
90
+ "jerk": 0.0016622950788587332,
91
+ "path_rotation": 0.06584077328443527,
92
+ "path_translation": 0.0010324910981580615,
93
+ "residual_magnitude": 0.0,
94
+ "total": 0.357170432806015,
95
+ "velocity": 0.053741686046123505
96
+ },
97
+ "metrics": {
98
+ "acceleration_rmse": 0.12457380443811417,
99
+ "command_acceleration_rms": 0.10650406777858734,
100
+ "command_jerk_rms": 2.6437199115753174,
101
+ "endpoint_rotation_rmse_deg": 3.9696099758148193,
102
+ "endpoint_translation_rmse_mm": 6.296663284301758,
103
+ "jerk_rmse": 3.105485677719116,
104
+ "path_rotation_rmse_deg": 2.5149660110473633,
105
+ "path_translation_rmse_mm": 3.9353995323181152,
106
+ "residual_rotation_mean_deg": 0.0,
107
+ "residual_translation_mean_mm": 0.0,
108
+ "velocity_rmse": 0.03739766776561737
109
+ },
110
+ "per_group_metrics": {
111
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
112
+ "acceleration_rmse": 0.12432977557182312,
113
+ "command_acceleration_rms": 0.11089512705802917,
114
+ "command_jerk_rms": 2.782877206802368,
115
+ "endpoint_rotation_rmse_deg": 4.350729465484619,
116
+ "endpoint_translation_rmse_mm": 6.224776744842529,
117
+ "jerk_rmse": 3.11281418800354,
118
+ "path_rotation_rmse_deg": 2.7672295570373535,
119
+ "path_translation_rmse_mm": 3.89414119720459,
120
+ "residual_rotation_mean_deg": 0.0,
121
+ "residual_translation_mean_mm": 0.0,
122
+ "velocity_rmse": 0.040785618126392365
123
+ },
124
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
125
+ "acceleration_rmse": 0.14718665182590485,
126
+ "command_acceleration_rms": 0.12139102071523666,
127
+ "command_jerk_rms": 2.9899542331695557,
128
+ "endpoint_rotation_rmse_deg": 4.108950138092041,
129
+ "endpoint_translation_rmse_mm": 7.592531204223633,
130
+ "jerk_rmse": 3.647094488143921,
131
+ "path_rotation_rmse_deg": 2.6102445125579834,
132
+ "path_translation_rmse_mm": 4.781116962432861,
133
+ "residual_rotation_mean_deg": 0.0,
134
+ "residual_translation_mean_mm": 0.0,
135
+ "velocity_rmse": 0.03910675644874573
136
+ },
137
+ "robotwin_generic::050_bell/base0::candidate_0018": {
138
+ "acceleration_rmse": 0.11602437496185303,
139
+ "command_acceleration_rms": 0.09816119074821472,
140
+ "command_jerk_rms": 2.3722805976867676,
141
+ "endpoint_rotation_rmse_deg": 3.9214015007019043,
142
+ "endpoint_translation_rmse_mm": 5.605432510375977,
143
+ "jerk_rmse": 2.9190683364868164,
144
+ "path_rotation_rmse_deg": 2.4512875080108643,
145
+ "path_translation_rmse_mm": 3.457277774810791,
146
+ "residual_rotation_mean_deg": 0.0,
147
+ "residual_translation_mean_mm": 0.0,
148
+ "velocity_rmse": 0.03643663600087166
149
+ },
150
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
151
+ "acceleration_rmse": 0.0005567761836573482,
152
+ "command_acceleration_rms": 0.00045566412154585123,
153
+ "command_jerk_rms": 0.017189867794513702,
154
+ "endpoint_rotation_rmse_deg": 5.807064056396484,
155
+ "endpoint_translation_rmse_mm": 6.1670756340026855,
156
+ "jerk_rmse": 0.01718975231051445,
157
+ "path_rotation_rmse_deg": 3.5094783306121826,
158
+ "path_translation_rmse_mm": 3.727048635482788,
159
+ "residual_rotation_mean_deg": 0.0,
160
+ "residual_translation_mean_mm": 0.0,
161
+ "velocity_rmse": 0.05181599035859108
162
+ },
163
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
164
+ "acceleration_rmse": 0.12440907955169678,
165
+ "command_acceleration_rms": 0.09963987022638321,
166
+ "command_jerk_rms": 2.5165276527404785,
167
+ "endpoint_rotation_rmse_deg": 2.6556525230407715,
168
+ "endpoint_translation_rmse_mm": 6.43054723739624,
169
+ "jerk_rmse": 3.1067402362823486,
170
+ "path_rotation_rmse_deg": 1.726789116859436,
171
+ "path_translation_rmse_mm": 4.088305950164795,
172
+ "residual_rotation_mean_deg": 0.0,
173
+ "residual_translation_mean_mm": 0.0,
174
+ "velocity_rmse": 0.02654867060482502
175
+ },
176
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
177
+ "acceleration_rmse": 0.13240712881088257,
178
+ "command_acceleration_rms": 0.08462731540203094,
179
+ "command_jerk_rms": 1.9978476762771606,
180
+ "endpoint_rotation_rmse_deg": 3.0848066806793213,
181
+ "endpoint_translation_rmse_mm": 5.198719024658203,
182
+ "jerk_rmse": 3.147235155105591,
183
+ "path_rotation_rmse_deg": 1.9724689722061157,
184
+ "path_translation_rmse_mm": 3.3021345138549805,
185
+ "residual_rotation_mean_deg": 0.0,
186
+ "residual_translation_mean_mm": 0.0,
187
+ "velocity_rmse": 0.030027437955141068
188
+ },
189
+ "robotwin_generic::071_can/base1::candidate_0084": {
190
+ "acceleration_rmse": 0.11730268597602844,
191
+ "command_acceleration_rms": 0.10736808180809021,
192
+ "command_jerk_rms": 2.672065019607544,
193
+ "endpoint_rotation_rmse_deg": 3.5667669773101807,
194
+ "endpoint_translation_rmse_mm": 6.416263580322266,
195
+ "jerk_rmse": 2.9105913639068604,
196
+ "path_rotation_rmse_deg": 2.2308413982391357,
197
+ "path_translation_rmse_mm": 3.9885590076446533,
198
+ "residual_rotation_mean_deg": 0.0,
199
+ "residual_translation_mean_mm": 0.0,
200
+ "velocity_rmse": 0.03380422294139862
201
+ },
202
+ "robotwin_generic::107_soap/base1::candidate_0106": {
203
+ "acceleration_rmse": 0.11784116178750992,
204
+ "command_acceleration_rms": 0.09079129248857498,
205
+ "command_jerk_rms": 2.2749669551849365,
206
+ "endpoint_rotation_rmse_deg": 3.645209312438965,
207
+ "endpoint_translation_rmse_mm": 5.499370574951172,
208
+ "jerk_rmse": 2.957634449005127,
209
+ "path_rotation_rmse_deg": 2.308201313018799,
210
+ "path_translation_rmse_mm": 3.419480800628662,
211
+ "residual_rotation_mean_deg": 0.0,
212
+ "residual_translation_mean_mm": 0.0,
213
+ "velocity_rmse": 0.0345214307308197
214
+ },
215
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
216
+ "acceleration_rmse": 0.11204928904771805,
217
+ "command_acceleration_rms": 0.09804097563028336,
218
+ "command_jerk_rms": 2.3640592098236084,
219
+ "endpoint_rotation_rmse_deg": 3.7972915172576904,
220
+ "endpoint_translation_rmse_mm": 6.545653343200684,
221
+ "jerk_rmse": 2.747910737991333,
222
+ "path_rotation_rmse_deg": 2.393160104751587,
223
+ "path_translation_rmse_mm": 4.025181770324707,
224
+ "residual_rotation_mean_deg": 0.0,
225
+ "residual_translation_mean_mm": 0.0,
226
+ "velocity_rmse": 0.035542167723178864
227
+ }
228
+ },
229
+ "prediction_source": "base_proposal",
230
+ "primary_path_score": {
231
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
232
+ "group_macro": 0.8735849380493164,
233
+ "micro": 0.8965331554412841
234
+ },
235
+ "run_id": "E002-R001",
236
+ "sample_count": 1298,
237
+ "schema_version": "precision.action_chunk.evaluation.v1",
238
+ "source_family_metrics": {
239
+ "rlbench": {
240
+ "group_count": 1,
241
+ "metrics": {
242
+ "acceleration_rmse": 0.12432977557182312,
243
+ "command_acceleration_rms": 0.11089512705802917,
244
+ "command_jerk_rms": 2.782877206802368,
245
+ "endpoint_rotation_rmse_deg": 4.350729465484619,
246
+ "endpoint_translation_rmse_mm": 6.224776744842529,
247
+ "jerk_rmse": 3.11281418800354,
248
+ "path_rotation_rmse_deg": 2.7672295570373535,
249
+ "path_translation_rmse_mm": 3.89414119720459,
250
+ "residual_rotation_mean_deg": 0.0,
251
+ "residual_translation_mean_mm": 0.0,
252
+ "velocity_rmse": 0.040785618126392365
253
+ },
254
+ "path_score": 0.9428600311279296,
255
+ "sample_count": 520
256
+ },
257
+ "robotwin_generic": {
258
+ "group_count": 8,
259
+ "metrics": {
260
+ "acceleration_rmse": 0.12473664432764053,
261
+ "command_acceleration_rms": 0.10346533358097076,
262
+ "command_jerk_rms": 2.546473503112793,
263
+ "endpoint_rotation_rmse_deg": 3.693011522293091,
264
+ "endpoint_translation_rmse_mm": 6.344256401062012,
265
+ "jerk_rmse": 3.1005780696868896,
266
+ "path_rotation_rmse_deg": 2.3311872482299805,
267
+ "path_translation_rmse_mm": 3.962736129760742,
268
+ "residual_rotation_mean_deg": 0.0,
269
+ "residual_translation_mean_mm": 0.0,
270
+ "velocity_rmse": 0.03495059534907341
271
+ },
272
+ "path_score": 0.8625110626220703,
273
+ "sample_count": 778
274
+ }
275
+ },
276
+ "split": "validation",
277
+ "test_firewall_opened": false
278
+ }
experiments/E002/runs/E002-R001/run_manifest.yaml ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R001
4
+ experiment_id: E002
5
+ condition_id: c-endpoint-5d2dff166d
6
+ status: completed
7
+ started_at: '2026-08-10T11:12:08.698434+00:00'
8
+ ended_at: '2026-08-10T11:18:50.394652+00:00'
9
+ condition:
10
+ model:
11
+ action_head: endpoint
12
+ base_conditioning: none
13
+ horizon_steps: 1
14
+ trajectory_basis: none
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R001-endpoint.yaml --data-root /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples
67
+ --output-dir /home/coder/share/experiment-runs/E002/E002-R001
68
+ metrics:
69
+ validation:
70
+ loss/total: 0.21126084798840786
71
+ loss/translation: 0.04712994648100599
72
+ loss/rotation_vector: 0.04120707774795994
73
+ loss/rotation_geodesic: 0.42234639320608647
74
+ loss/joint: 0.03408178494804078
75
+ loss/gripper: 0.0029633291347672696
76
+ metric/translation_mm: 23.808171852711354
77
+ metric/rotation_deg: 12.669311751202553
78
+ metric/joint_mae_rad: 0.08809902482251357
79
+ metric/gripper_accuracy: 1.0
80
+ checkpoint:
81
+ path: /home/coder/share/experiment-runs/E002/E002-R001/best.pt
82
+ sha256: 0126a82280cb7db4b733f63ab3cae79d47360c9bea700034e4c2a07b646c94da
83
+ selection_metric: loss/total
84
+ selection_split: validation
85
+ artifacts:
86
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R001/metrics.jsonl
87
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R001/evaluation_validation.json
88
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R001/condition.yaml
89
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R001/config_resolved.yaml
90
+ compute:
91
+ wall_seconds: 401.5836740859959
92
+ gpu_hours: 0.1115510205794433
93
+ peak_memory_bytes: 2885576192
experiments/E002/runs/E002-R002/condition_config.yaml ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.condition_config.v1
2
+ experiment_id: E002
3
+ run_id: E002-R002
4
+ condition_id: c-residual-h16-f186e90df2
5
+ condition_summary: residual-h16
6
+ condition:
7
+ model:
8
+ action_head: residual_chunk
9
+ base_conditioning: E002-R001-profiled-proposal
10
+ horizon_steps: 16
11
+ trajectory_basis: independent_h16
12
+ resolved_config:
13
+ data:
14
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
15
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
16
+ variant_id: validated-physical-recovery-15k-v1
17
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
18
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
19
+ sample_count: 15000
20
+ eligible_sample_count: 11628
21
+ eligible_group_count: 89
22
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
23
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
24
+ split_namespace: 'E002-source-group-split-v1::'
25
+ validation_fraction: 0.1
26
+ test_fraction: 0.1
27
+ geometry:
28
+ cumulative_waypoint_representation: current_tool_relative_pose
29
+ step_representation: integrated_body_twist
30
+ step_integration: right_se3_exponential
31
+ residual_curve_frame: initial_local_tool_frame
32
+ residual_composition: base_right_multiply_se3_exponential
33
+ structured_translation: se3_tangent_translation_control_points_m
34
+ structured_rotation: so3_lie_algebra_control_points_rad
35
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
36
+ residual_boundary_at_t0: identity
37
+ model:
38
+ backbone:
39
+ hidden_size: 256
40
+ fusion_layers: 4
41
+ fusion_heads: 8
42
+ fusion_mlp_ratio: 4
43
+ dropout: 0.1
44
+ vision_backend: conv
45
+ text_backend: embedding
46
+ pretrained_model_name: google/siglip-base-patch16-224
47
+ freeze_pretrained: true
48
+ image_size: 224
49
+ vision_tokens: 64
50
+ depth_tokens: 36
51
+ point_tokens: 32
52
+ point_samples: 256
53
+ max_text_tokens: 32
54
+ text_vocab_size: 32768
55
+ max_joints: 7
56
+ aloha_joints: 6
57
+ panda_joints: 7
58
+ translation_scale_m: 0.05
59
+ rotation_scale_rad: 0.5235987755982988
60
+ joint_scale_rad: 0.5
61
+ horizon_steps: 16
62
+ horizon_seconds: 0.8
63
+ decoder_layers: 4
64
+ decoder_heads: 8
65
+ decoder_mlp_ratio: 4
66
+ control_points: 6
67
+ maximum_step_translation_m: 0.01
68
+ maximum_step_rotation_rad: 0.2
69
+ maximum_residual_translation_m: 0.05
70
+ maximum_residual_rotation_rad: 0.52
71
+ action_head: residual_chunk
72
+ base_conditioning: E002-R001-profiled-proposal
73
+ trajectory_basis: independent_h16
74
+ loss:
75
+ path_translation: 1.0
76
+ path_rotation: 1.0
77
+ endpoint_translation: 2.0
78
+ endpoint_rotation: 2.0
79
+ velocity: 1.0
80
+ acceleration: 0.1
81
+ jerk: 0.05
82
+ residual_magnitude: 0.02
83
+ training:
84
+ seed: 42
85
+ batch_size: 64
86
+ gradient_accumulation_steps: 1
87
+ epochs: 100
88
+ max_steps: 3000
89
+ learning_rate: 0.0003
90
+ weight_decay: 0.05
91
+ warmup_fraction: 0.05
92
+ gradient_clip_norm: 1.0
93
+ precision: bf16
94
+ workers: 8
95
+ log_every_steps: 10
96
+ validation_every_steps: 250
97
+ checkpoint_every_steps: 500
98
+ evaluation:
99
+ version: se3-spline-eval-v1
100
+ trajectory_loss:
101
+ path_translation: 1.0
102
+ path_rotation: 1.0
103
+ endpoint_translation: 2.0
104
+ endpoint_rotation: 2.0
105
+ velocity: 1.0
106
+ acceleration: 0.1
107
+ jerk: 0.05
108
+ residual_magnitude: 0.02
109
+ selection_split: validation
110
+ test_firewall_opened: false
111
+ group_bootstrap_seed: 20260810
112
+ group_bootstrap_resamples: 10000
experiments/E002/runs/E002-R002/evaluation_validation.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R002/best.pt",
8
+ "sha256": "c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-residual-h16-f186e90df2",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R002/run_manifest.yaml",
15
+ "manifest_sha256": "6e4e3a93e8e2a046a47309d8f904abbedcc01ca2f577b837e214106a9d068e73",
16
+ "run_id": "E002-R002"
17
+ },
18
+ "condition_id": "c-residual-h16-f186e90df2",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.12913274992671278,
30
+ "upper": 0.15617000725534227
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.11790996123519208,
34
+ "upper": 0.13553185007638402
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 2.875136958228217,
38
+ "upper": 3.3160451802942488
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.29721736907959,
42
+ "upper": 4.666348218917846
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.61740255355835,
46
+ "upper": 6.495758643415239
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7692517068650987,
50
+ "upper": 0.9713847054375542
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 3.155406200223499,
54
+ "upper": 3.7974631309509275
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 2.0693869060940213,
58
+ "upper": 2.87911589940389
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.491334999932183,
62
+ "upper": 4.021194886498981
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.6610424452357822,
66
+ "upper": 0.872592170205381
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.6254603839582867,
70
+ "upper": 0.7778024170133803
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.03161394658188025,
74
+ "upper": 0.043074266581485665
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.144492506980896,
84
+ "command_acceleration_rms": 0.12820853292942047,
85
+ "command_jerk_rms": 3.1372432708740234,
86
+ "endpoint_rotation_rmse_deg": 3.8582987785339355,
87
+ "endpoint_translation_rmse_mm": 6.0623393058776855,
88
+ "jerk_rmse": 3.533108949661255,
89
+ "path_rotation_rmse_deg": 2.405137062072754,
90
+ "path_translation_rmse_mm": 3.7594053745269775,
91
+ "residual_rotation_mean_deg": 0.7692350149154663,
92
+ "residual_translation_mean_mm": 0.6971892714500427,
93
+ "velocity_rmse": 0.03634699434041977
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.110406999999999,
98
+ "p95_ms": 8.200868949999998,
99
+ "p99_ms": 10.43797239,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0015241311630234122,
105
+ "endpoint_rotation": 0.10943228751420975,
106
+ "endpoint_translation": 0.002436618087813258,
107
+ "jerk": 0.00226704147644341,
108
+ "path_rotation": 0.061600614339113235,
109
+ "path_translation": 0.0009408381883986294,
110
+ "residual_magnitude": 0.03566799312829971,
111
+ "total": 0.33692434430122375,
112
+ "velocity": 0.04966599866747856
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.1495567262172699,
116
+ "command_acceleration_rms": 0.13441726565361023,
117
+ "command_jerk_rms": 3.3160297870635986,
118
+ "endpoint_rotation_rmse_deg": 3.789569139480591,
119
+ "endpoint_translation_rmse_mm": 6.045598983764648,
120
+ "jerk_rmse": 3.691244125366211,
121
+ "path_rotation_rmse_deg": 2.380246877670288,
122
+ "path_translation_rmse_mm": 3.7566702365875244,
123
+ "residual_rotation_mean_deg": 0.672172486782074,
124
+ "residual_translation_mean_mm": 0.663112223148346,
125
+ "velocity_rmse": 0.03584917262196541
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.15038879215717316,
130
+ "command_acceleration_rms": 0.1392577588558197,
131
+ "command_jerk_rms": 3.4107391834259033,
132
+ "endpoint_rotation_rmse_deg": 4.199632167816162,
133
+ "endpoint_translation_rmse_mm": 6.008756637573242,
134
+ "jerk_rmse": 3.691063404083252,
135
+ "path_rotation_rmse_deg": 2.6454405784606934,
136
+ "path_translation_rmse_mm": 3.743166446685791,
137
+ "residual_rotation_mean_deg": 0.45984548330307007,
138
+ "residual_translation_mean_mm": 0.5546795725822449,
139
+ "velocity_rmse": 0.039565857499837875
140
+ },
141
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
142
+ "acceleration_rmse": 0.15587416291236877,
143
+ "command_acceleration_rms": 0.13095782697200775,
144
+ "command_jerk_rms": 3.4089276790618896,
145
+ "endpoint_rotation_rmse_deg": 3.9784700870513916,
146
+ "endpoint_translation_rmse_mm": 6.946713924407959,
147
+ "jerk_rmse": 3.987175464630127,
148
+ "path_rotation_rmse_deg": 2.506291389465332,
149
+ "path_translation_rmse_mm": 4.335691452026367,
150
+ "residual_rotation_mean_deg": 0.7182945013046265,
151
+ "residual_translation_mean_mm": 0.9254800081253052,
152
+ "velocity_rmse": 0.03756774961948395
153
+ },
154
+ "robotwin_generic::050_bell/base0::candidate_0018": {
155
+ "acceleration_rmse": 0.1444813460111618,
156
+ "command_acceleration_rms": 0.13011473417282104,
157
+ "command_jerk_rms": 3.1705923080444336,
158
+ "endpoint_rotation_rmse_deg": 3.414102554321289,
159
+ "endpoint_translation_rmse_mm": 5.582004547119141,
160
+ "jerk_rmse": 3.5780766010284424,
161
+ "path_rotation_rmse_deg": 2.1180219650268555,
162
+ "path_translation_rmse_mm": 3.4302897453308105,
163
+ "residual_rotation_mean_deg": 0.8319661617279053,
164
+ "residual_translation_mean_mm": 0.7032113075256348,
165
+ "velocity_rmse": 0.03218117728829384
166
+ },
167
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
168
+ "acceleration_rmse": 0.09131758660078049,
169
+ "command_acceleration_rms": 0.09131619334220886,
170
+ "command_jerk_rms": 2.181763172149658,
171
+ "endpoint_rotation_rmse_deg": 6.780344009399414,
172
+ "endpoint_translation_rmse_mm": 7.127496242523193,
173
+ "jerk_rmse": 2.1817626953125,
174
+ "path_rotation_rmse_deg": 4.104794025421143,
175
+ "path_translation_rmse_mm": 4.363109111785889,
176
+ "residual_rotation_mean_deg": 0.5674796104431152,
177
+ "residual_translation_mean_mm": 0.6312280893325806,
178
+ "velocity_rmse": 0.06063513085246086
179
+ },
180
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
181
+ "acceleration_rmse": 0.14704690873622894,
182
+ "command_acceleration_rms": 0.12569846212863922,
183
+ "command_jerk_rms": 3.1258838176727295,
184
+ "endpoint_rotation_rmse_deg": 2.983164072036743,
185
+ "endpoint_translation_rmse_mm": 6.01630973815918,
186
+ "jerk_rmse": 3.631091594696045,
187
+ "path_rotation_rmse_deg": 1.887567400932312,
188
+ "path_translation_rmse_mm": 3.7935662269592285,
189
+ "residual_rotation_mean_deg": 0.9006107449531555,
190
+ "residual_translation_mean_mm": 0.6452526450157166,
191
+ "velocity_rmse": 0.02905953861773014
192
+ },
193
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
194
+ "acceleration_rmse": 0.1694936454296112,
195
+ "command_acceleration_rms": 0.13346831500530243,
196
+ "command_jerk_rms": 3.077533483505249,
197
+ "endpoint_rotation_rmse_deg": 2.9263644218444824,
198
+ "endpoint_translation_rmse_mm": 4.760553359985352,
199
+ "jerk_rmse": 3.9148061275482178,
200
+ "path_rotation_rmse_deg": 1.850063681602478,
201
+ "path_translation_rmse_mm": 2.988938808441162,
202
+ "residual_rotation_mean_deg": 0.9843810796737671,
203
+ "residual_translation_mean_mm": 0.8447241187095642,
204
+ "velocity_rmse": 0.029001181945204735
205
+ },
206
+ "robotwin_generic::071_can/base1::candidate_0084": {
207
+ "acceleration_rmse": 0.13344013690948486,
208
+ "command_acceleration_rms": 0.12478600442409515,
209
+ "command_jerk_rms": 3.220496892929077,
210
+ "endpoint_rotation_rmse_deg": 3.3317511081695557,
211
+ "endpoint_translation_rmse_mm": 6.138230323791504,
212
+ "jerk_rmse": 3.400482416152954,
213
+ "path_rotation_rmse_deg": 2.0716307163238525,
214
+ "path_translation_rmse_mm": 3.78041410446167,
215
+ "residual_rotation_mean_deg": 0.7835702300071716,
216
+ "residual_translation_mean_mm": 0.7476160526275635,
217
+ "velocity_rmse": 0.03152817487716675
218
+ },
219
+ "robotwin_generic::107_soap/base1::candidate_0106": {
220
+ "acceleration_rmse": 0.15871137380599976,
221
+ "command_acceleration_rms": 0.13939915597438812,
222
+ "command_jerk_rms": 3.347158670425415,
223
+ "endpoint_rotation_rmse_deg": 3.444718837738037,
224
+ "endpoint_translation_rmse_mm": 5.598421573638916,
225
+ "jerk_rmse": 3.837700605392456,
226
+ "path_rotation_rmse_deg": 2.184645652770996,
227
+ "path_translation_rmse_mm": 3.4819819927215576,
228
+ "residual_rotation_mean_deg": 0.7184168696403503,
229
+ "residual_translation_mean_mm": 0.6539862155914307,
230
+ "velocity_rmse": 0.032942239195108414
231
+ },
232
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
233
+ "acceleration_rmse": 0.14967872202396393,
234
+ "command_acceleration_rms": 0.13887833058834076,
235
+ "command_jerk_rms": 3.2920944690704346,
236
+ "endpoint_rotation_rmse_deg": 3.6661412715911865,
237
+ "endpoint_translation_rmse_mm": 6.382567405700684,
238
+ "jerk_rmse": 3.575821876525879,
239
+ "path_rotation_rmse_deg": 2.2777774333953857,
240
+ "path_translation_rmse_mm": 3.9174914360046387,
241
+ "residual_rotation_mean_deg": 0.9585511088371277,
242
+ "residual_translation_mean_mm": 0.5685251951217651,
243
+ "velocity_rmse": 0.03464192524552345
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8569679498672484,
250
+ "micro": 0.8517163991928101
251
+ },
252
+ "run_id": "E002-R002",
253
+ "sample_count": 1298,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.15038879215717316,
260
+ "command_acceleration_rms": 0.1392577588558197,
261
+ "command_jerk_rms": 3.4107391834259033,
262
+ "endpoint_rotation_rmse_deg": 4.199632167816162,
263
+ "endpoint_translation_rmse_mm": 6.008756637573242,
264
+ "jerk_rmse": 3.691063404083252,
265
+ "path_rotation_rmse_deg": 2.6454405784606934,
266
+ "path_translation_rmse_mm": 3.743166446685791,
267
+ "residual_rotation_mean_deg": 0.45984548330307007,
268
+ "residual_translation_mean_mm": 0.5546795725822449,
269
+ "velocity_rmse": 0.039565857499837875
270
+ },
271
+ "path_score": 0.9034047603607178,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.14899800717830658,
278
+ "command_acceleration_rms": 0.13108235597610474,
279
+ "command_jerk_rms": 3.251190185546875,
280
+ "endpoint_rotation_rmse_deg": 3.488720178604126,
281
+ "endpoint_translation_rmse_mm": 6.070099353790283,
282
+ "jerk_rmse": 3.6913645267486572,
283
+ "path_rotation_rmse_deg": 2.1851251125335693,
284
+ "path_translation_rmse_mm": 3.765669107437134,
285
+ "residual_rotation_mean_deg": 0.8140876889228821,
286
+ "residual_translation_mean_mm": 0.7355865240097046,
287
+ "velocity_rmse": 0.0331333726644516
288
+ },
289
+ "path_score": 0.8135919332504273,
290
+ "sample_count": 778
291
+ }
292
+ },
293
+ "split": "validation",
294
+ "test_firewall_opened": false
295
+ }
experiments/E002/runs/E002-R002/run_manifest.yaml ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R002
4
+ experiment_id: E002
5
+ condition_id: c-residual-h16-f186e90df2
6
+ status: completed
7
+ started_at: '2026-08-10T11:31:58.411131+00:00'
8
+ ended_at: '2026-08-10T11:37:02.967653+00:00'
9
+ condition:
10
+ model:
11
+ action_head: residual_chunk
12
+ base_conditioning: E002-R001-profiled-proposal
13
+ horizon_steps: 16
14
+ trajectory_basis: independent_h16
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R002-residual-h16.yaml --data-root
67
+ /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir
68
+ /home/coder/share/experiment-runs/E002/E002-R002 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
69
+ metrics:
70
+ validation:
71
+ loss/total: 0.33688340996189736
72
+ loss/path_translation: 0.000940837494032357
73
+ loss/path_rotation: 0.0615845355097612
74
+ loss/endpoint_translation: 0.0024366182872798806
75
+ loss/endpoint_rotation: 0.10941956785775847
76
+ loss/velocity: 0.04966655885011463
77
+ loss/acceleration: 0.0015241503069759884
78
+ loss/jerk: 0.0022666104369651876
79
+ loss/residual_magnitude: 0.0356679620689071
80
+ metric/path_translation_rmse_mm: 3.753689542573479
81
+ metric/path_rotation_rmse_deg: 2.3661541755100246
82
+ metric/endpoint_translation_rmse_mm: 6.03998319761044
83
+ metric/endpoint_rotation_rmse_deg: 3.7676057393084323
84
+ metric/velocity_rmse: 0.03566566113101315
85
+ metric/acceleration_rmse: 0.1493868921497386
86
+ metric/jerk_rmse: 3.6834481236379943
87
+ metric/command_acceleration_rms: 0.13423365026207293
88
+ metric/command_jerk_rms: 3.309643663500417
89
+ metric/residual_translation_mean_mm: 0.66311021459121
90
+ metric/residual_rotation_mean_deg: 0.6721727473342731
91
+ checkpoint:
92
+ path: /home/coder/share/experiment-runs/E002/E002-R002/best.pt
93
+ sha256: c7cd3aac047e4ed726f086b292bd74121f513484e9979c773cafc606a90d8da9
94
+ selection_metric: loss/total
95
+ selection_split: validation
96
+ artifacts:
97
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R002/metrics.jsonl
98
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R002/evaluation_validation.json
99
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R002/condition.yaml
100
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R002/config_resolved.yaml
101
+ compute:
102
+ wall_seconds: 304.37659596299636
103
+ gpu_hours: 0.08454905443416566
104
+ peak_memory_bytes: 2952011264
experiments/E002/runs/E002-R003/condition_config.yaml ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.condition_config.v1
2
+ experiment_id: E002
3
+ run_id: E002-R003
4
+ condition_id: c-cubic-se3-cp6-43b928e718
5
+ condition_summary: cubic-se3-cp6
6
+ condition:
7
+ model:
8
+ action_head: cubic_bspline
9
+ base_conditioning: E002-R001-profiled-proposal
10
+ horizon_steps: 16
11
+ trajectory_basis: clamped_cubic_bspline_degree3_cp6
12
+ resolved_config:
13
+ data:
14
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
15
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
16
+ variant_id: validated-physical-recovery-15k-v1
17
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
18
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
19
+ sample_count: 15000
20
+ eligible_sample_count: 11628
21
+ eligible_group_count: 89
22
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
23
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
24
+ split_namespace: 'E002-source-group-split-v1::'
25
+ validation_fraction: 0.1
26
+ test_fraction: 0.1
27
+ geometry:
28
+ cumulative_waypoint_representation: current_tool_relative_pose
29
+ step_representation: integrated_body_twist
30
+ step_integration: right_se3_exponential
31
+ residual_curve_frame: initial_local_tool_frame
32
+ residual_composition: base_right_multiply_se3_exponential
33
+ structured_translation: se3_tangent_translation_control_points_m
34
+ structured_rotation: so3_lie_algebra_control_points_rad
35
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
36
+ residual_boundary_at_t0: identity
37
+ model:
38
+ backbone:
39
+ hidden_size: 256
40
+ fusion_layers: 4
41
+ fusion_heads: 8
42
+ fusion_mlp_ratio: 4
43
+ dropout: 0.1
44
+ vision_backend: conv
45
+ text_backend: embedding
46
+ pretrained_model_name: google/siglip-base-patch16-224
47
+ freeze_pretrained: true
48
+ image_size: 224
49
+ vision_tokens: 64
50
+ depth_tokens: 36
51
+ point_tokens: 32
52
+ point_samples: 256
53
+ max_text_tokens: 32
54
+ text_vocab_size: 32768
55
+ max_joints: 7
56
+ aloha_joints: 6
57
+ panda_joints: 7
58
+ translation_scale_m: 0.05
59
+ rotation_scale_rad: 0.5235987755982988
60
+ joint_scale_rad: 0.5
61
+ horizon_steps: 16
62
+ horizon_seconds: 0.8
63
+ decoder_layers: 4
64
+ decoder_heads: 8
65
+ decoder_mlp_ratio: 4
66
+ control_points: 6
67
+ maximum_step_translation_m: 0.01
68
+ maximum_step_rotation_rad: 0.2
69
+ maximum_residual_translation_m: 0.05
70
+ maximum_residual_rotation_rad: 0.52
71
+ action_head: cubic_bspline
72
+ base_conditioning: E002-R001-profiled-proposal
73
+ trajectory_basis: clamped_cubic_bspline_degree3_cp6
74
+ loss:
75
+ path_translation: 1.0
76
+ path_rotation: 1.0
77
+ endpoint_translation: 2.0
78
+ endpoint_rotation: 2.0
79
+ velocity: 1.0
80
+ acceleration: 0.1
81
+ jerk: 0.05
82
+ residual_magnitude: 0.02
83
+ training:
84
+ seed: 42
85
+ batch_size: 64
86
+ gradient_accumulation_steps: 1
87
+ epochs: 100
88
+ max_steps: 3000
89
+ learning_rate: 0.0003
90
+ weight_decay: 0.05
91
+ warmup_fraction: 0.05
92
+ gradient_clip_norm: 1.0
93
+ precision: bf16
94
+ workers: 8
95
+ log_every_steps: 10
96
+ validation_every_steps: 250
97
+ checkpoint_every_steps: 500
98
+ evaluation:
99
+ version: se3-spline-eval-v1
100
+ trajectory_loss:
101
+ path_translation: 1.0
102
+ path_rotation: 1.0
103
+ endpoint_translation: 2.0
104
+ endpoint_rotation: 2.0
105
+ velocity: 1.0
106
+ acceleration: 0.1
107
+ jerk: 0.05
108
+ residual_magnitude: 0.02
109
+ selection_split: validation
110
+ test_firewall_opened: false
111
+ group_bootstrap_seed: 20260810
112
+ group_bootstrap_resamples: 10000
experiments/E002/runs/E002-R003/evaluation_validation.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R003/best.pt",
8
+ "sha256": "3490f23a0a5bcddb9d1998d08af37318023c8f792b17c68c1562b317dc88554a"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-cubic-se3-cp6-43b928e718",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R003/run_manifest.yaml",
15
+ "manifest_sha256": "6ec30c7b61750729b9750a0fbb86e983c1192defd433e9986c69b6be3fb77298",
16
+ "run_id": "E002-R003"
17
+ },
18
+ "condition_id": "c-cubic-se3-cp6-43b928e718",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.11298756880892648,
30
+ "upper": 0.1394385862888561
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.09995081027348836,
34
+ "upper": 0.11602214109152555
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 1.8295588115851085,
38
+ "upper": 2.6620728174845376
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.2371414442857107,
42
+ "upper": 4.614357245630688
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.5499372945891485,
46
+ "upper": 6.477358905474345
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7578355182541741,
50
+ "upper": 0.9678520631127886
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 2.2059771882163153,
54
+ "upper": 3.228559059566922
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 2.0445298645231462,
58
+ "upper": 2.8757890260881847
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.434231901168823,
62
+ "upper": 3.9773682713508607
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.7555553118387858,
66
+ "upper": 0.9298186732663049
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.8847316967116462,
70
+ "upper": 1.1563906035489506
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.031309114784623185,
74
+ "upper": 0.042778491559955806
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.1268954873085022,
84
+ "command_acceleration_rms": 0.10837496817111969,
85
+ "command_jerk_rms": 2.313187599182129,
86
+ "endpoint_rotation_rmse_deg": 3.8112881183624268,
87
+ "endpoint_translation_rmse_mm": 6.002523422241211,
88
+ "jerk_rmse": 2.8165056705474854,
89
+ "path_rotation_rmse_deg": 2.38803768157959,
90
+ "path_translation_rmse_mm": 3.7016143798828125,
91
+ "residual_rotation_mean_deg": 0.84266597032547,
92
+ "residual_translation_mean_mm": 1.0173211097717285,
93
+ "velocity_rmse": 0.03602871671319008
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.1211475,
98
+ "p95_ms": 8.21552085,
99
+ "p99_ms": 8.44476369,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0012071238597854972,
105
+ "endpoint_rotation": 0.10867936909198761,
106
+ "endpoint_translation": 0.002355183707550168,
107
+ "jerk": 0.0016878115711733699,
108
+ "path_rotation": 0.06108936667442322,
109
+ "path_translation": 0.0008991192444227636,
110
+ "residual_magnitude": 0.047036848962306976,
111
+ "total": 0.3334701955318451,
112
+ "velocity": 0.048266757279634476
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.12969663739204407,
116
+ "command_acceleration_rms": 0.11209042370319366,
117
+ "command_jerk_rms": 2.6622745990753174,
118
+ "endpoint_rotation_rmse_deg": 3.772080898284912,
119
+ "endpoint_translation_rmse_mm": 5.9437150955200195,
120
+ "jerk_rmse": 3.121913433074951,
121
+ "path_rotation_rmse_deg": 2.360351085662842,
122
+ "path_translation_rmse_mm": 3.672436237335205,
123
+ "residual_rotation_mean_deg": 0.7992115616798401,
124
+ "residual_translation_mean_mm": 1.0198233127593994,
125
+ "velocity_rmse": 0.03550861403346062
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.12533091008663177,
130
+ "command_acceleration_rms": 0.1117461621761322,
131
+ "command_jerk_rms": 2.788883924484253,
132
+ "endpoint_rotation_rmse_deg": 4.224048614501953,
133
+ "endpoint_translation_rmse_mm": 5.911834716796875,
134
+ "jerk_rmse": 3.118834972381592,
135
+ "path_rotation_rmse_deg": 2.6422619819641113,
136
+ "path_translation_rmse_mm": 3.6519975662231445,
137
+ "residual_rotation_mean_deg": 0.6977745294570923,
138
+ "residual_translation_mean_mm": 0.9536401033401489,
139
+ "velocity_rmse": 0.03928665071725845
140
+ },
141
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
142
+ "acceleration_rmse": 0.15442746877670288,
143
+ "command_acceleration_rms": 0.12963150441646576,
144
+ "command_jerk_rms": 3.0172955989837646,
145
+ "endpoint_rotation_rmse_deg": 3.862656831741333,
146
+ "endpoint_translation_rmse_mm": 6.788612365722656,
147
+ "jerk_rmse": 3.6710264682769775,
148
+ "path_rotation_rmse_deg": 2.423619270324707,
149
+ "path_translation_rmse_mm": 4.2186384201049805,
150
+ "residual_rotation_mean_deg": 0.9736085534095764,
151
+ "residual_translation_mean_mm": 1.314324975013733,
152
+ "velocity_rmse": 0.03680943325161934
153
+ },
154
+ "robotwin_generic::050_bell/base0::candidate_0018": {
155
+ "acceleration_rmse": 0.1225070133805275,
156
+ "command_acceleration_rms": 0.10519914329051971,
157
+ "command_jerk_rms": 2.399007558822632,
158
+ "endpoint_rotation_rmse_deg": 3.3820717334747314,
159
+ "endpoint_translation_rmse_mm": 5.527786731719971,
160
+ "jerk_rmse": 2.940819501876831,
161
+ "path_rotation_rmse_deg": 2.089040756225586,
162
+ "path_translation_rmse_mm": 3.3952102661132812,
163
+ "residual_rotation_mean_deg": 0.8335265517234802,
164
+ "residual_translation_mean_mm": 0.9504472613334656,
165
+ "velocity_rmse": 0.031815849244594574
166
+ },
167
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
168
+ "acceleration_rmse": 0.08224337548017502,
169
+ "command_acceleration_rms": 0.08224129676818848,
170
+ "command_jerk_rms": 0.6084536910057068,
171
+ "endpoint_rotation_rmse_deg": 6.687165260314941,
172
+ "endpoint_translation_rmse_mm": 7.383237838745117,
173
+ "jerk_rmse": 0.6084533929824829,
174
+ "path_rotation_rmse_deg": 4.139664649963379,
175
+ "path_translation_rmse_mm": 4.465749740600586,
176
+ "residual_rotation_mean_deg": 0.6802061200141907,
177
+ "residual_translation_mean_mm": 0.723755955696106,
178
+ "velocity_rmse": 0.060256969183683395
179
+ },
180
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
181
+ "acceleration_rmse": 0.13598355650901794,
182
+ "command_acceleration_rms": 0.11334140598773956,
183
+ "command_jerk_rms": 2.5503909587860107,
184
+ "endpoint_rotation_rmse_deg": 2.79191255569458,
185
+ "endpoint_translation_rmse_mm": 5.887653827667236,
186
+ "jerk_rmse": 3.1343507766723633,
187
+ "path_rotation_rmse_deg": 1.8062597513198853,
188
+ "path_translation_rmse_mm": 3.6813766956329346,
189
+ "residual_rotation_mean_deg": 0.9098313450813293,
190
+ "residual_translation_mean_mm": 1.0118743181228638,
191
+ "velocity_rmse": 0.02775019034743309
192
+ },
193
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
194
+ "acceleration_rmse": 0.15385229885578156,
195
+ "command_acceleration_rms": 0.11339045315980911,
196
+ "command_jerk_rms": 2.062548875808716,
197
+ "endpoint_rotation_rmse_deg": 2.887746572494507,
198
+ "endpoint_translation_rmse_mm": 4.715236186981201,
199
+ "jerk_rmse": 3.185887098312378,
200
+ "path_rotation_rmse_deg": 1.8494901657104492,
201
+ "path_translation_rmse_mm": 2.9501864910125732,
202
+ "residual_rotation_mean_deg": 1.072643518447876,
203
+ "residual_translation_mean_mm": 1.3127667903900146,
204
+ "velocity_rmse": 0.029317552223801613
205
+ },
206
+ "robotwin_generic::071_can/base1::candidate_0084": {
207
+ "acceleration_rmse": 0.11820736527442932,
208
+ "command_acceleration_rms": 0.10775069892406464,
209
+ "command_jerk_rms": 2.682610034942627,
210
+ "endpoint_rotation_rmse_deg": 3.3814568519592285,
211
+ "endpoint_translation_rmse_mm": 6.007859230041504,
212
+ "jerk_rmse": 2.919471502304077,
213
+ "path_rotation_rmse_deg": 2.0913233757019043,
214
+ "path_translation_rmse_mm": 3.676405906677246,
215
+ "residual_rotation_mean_deg": 0.8032221794128418,
216
+ "residual_translation_mean_mm": 1.2335833311080933,
217
+ "velocity_rmse": 0.0318651907145977
218
+ },
219
+ "robotwin_generic::107_soap/base1::candidate_0106": {
220
+ "acceleration_rmse": 0.12335247546434402,
221
+ "command_acceleration_rms": 0.09845532476902008,
222
+ "command_jerk_rms": 2.2949612140655518,
223
+ "endpoint_rotation_rmse_deg": 3.4214773178100586,
224
+ "endpoint_translation_rmse_mm": 5.568272113800049,
225
+ "jerk_rmse": 2.9741382598876953,
226
+ "path_rotation_rmse_deg": 2.1538639068603516,
227
+ "path_translation_rmse_mm": 3.451195001602173,
228
+ "residual_rotation_mean_deg": 0.6639298796653748,
229
+ "residual_translation_mean_mm": 0.7777648568153381,
230
+ "velocity_rmse": 0.03259320184588432
231
+ },
232
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
233
+ "acceleration_rmse": 0.12615501880645752,
234
+ "command_acceleration_rms": 0.11361874639987946,
235
+ "command_jerk_rms": 2.4145374298095703,
236
+ "endpoint_rotation_rmse_deg": 3.6630587577819824,
237
+ "endpoint_translation_rmse_mm": 6.232222557067871,
238
+ "jerk_rmse": 2.7955682277679443,
239
+ "path_rotation_rmse_deg": 2.2968173027038574,
240
+ "path_translation_rmse_mm": 3.82377028465271,
241
+ "residual_rotation_mean_deg": 0.9492508172988892,
242
+ "residual_translation_mean_mm": 0.8777318000793457,
243
+ "velocity_rmse": 0.034563399851322174
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8477689743041992,
250
+ "micro": 0.8393138408660888
251
+ },
252
+ "run_id": "E002-R003",
253
+ "sample_count": 1298,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.12533091008663177,
260
+ "command_acceleration_rms": 0.1117461621761322,
261
+ "command_jerk_rms": 2.788883924484253,
262
+ "endpoint_rotation_rmse_deg": 4.224048614501953,
263
+ "endpoint_translation_rmse_mm": 5.911834716796875,
264
+ "jerk_rmse": 3.118834972381592,
265
+ "path_rotation_rmse_deg": 2.6422619819641113,
266
+ "path_translation_rmse_mm": 3.6519975662231445,
267
+ "residual_rotation_mean_deg": 0.6977745294570923,
268
+ "residual_translation_mean_mm": 0.9536401033401489,
269
+ "velocity_rmse": 0.03928665071725845
270
+ },
271
+ "path_score": 0.8936521530151367,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.1325344294309616,
278
+ "command_acceleration_rms": 0.11231992393732071,
279
+ "command_jerk_rms": 2.5741817951202393,
280
+ "endpoint_rotation_rmse_deg": 3.4370152950286865,
281
+ "endpoint_translation_rmse_mm": 5.964929103851318,
282
+ "jerk_rmse": 3.123969316482544,
283
+ "path_rotation_rmse_deg": 2.1514289379119873,
284
+ "path_translation_rmse_mm": 3.6860344409942627,
285
+ "residual_rotation_mean_deg": 0.8670100569725037,
286
+ "residual_translation_mean_mm": 1.064058780670166,
287
+ "velocity_rmse": 0.03274128586053848
288
+ },
289
+ "path_score": 0.7988892316818237,
290
+ "sample_count": 778
291
+ }
292
+ },
293
+ "split": "validation",
294
+ "test_firewall_opened": false
295
+ }
experiments/E002/runs/E002-R003/run_manifest.yaml ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R003
4
+ experiment_id: E002
5
+ condition_id: c-cubic-se3-cp6-43b928e718
6
+ status: completed
7
+ started_at: '2026-08-10T11:37:25.747422+00:00'
8
+ ended_at: '2026-08-10T11:42:15.419706+00:00'
9
+ condition:
10
+ model:
11
+ action_head: cubic_bspline
12
+ base_conditioning: E002-R001-profiled-proposal
13
+ horizon_steps: 16
14
+ trajectory_basis: clamped_cubic_bspline_degree3_cp6
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R003-cubic-se3-cp6.yaml --data-root
67
+ /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir
68
+ /home/coder/share/experiment-runs/E002/E002-R003 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
69
+ metrics:
70
+ validation:
71
+ loss/total: 0.3333665318959299
72
+ loss/path_translation: 0.0008991088694677571
73
+ loss/path_rotation: 0.0610695354904526
74
+ loss/endpoint_translation: 0.002355156106176931
75
+ loss/endpoint_rotation: 0.1086380337529998
76
+ loss/velocity: 0.048265731894914836
77
+ loss/acceleration: 0.0012069116290931365
78
+ loss/jerk: 0.001686878622371778
79
+ loss/residual_magnitude: 0.04703686542476087
80
+ metric/path_translation_rmse_mm: 3.6693616519540044
81
+ metric/path_rotation_rmse_deg: 2.3440397380864124
82
+ metric/endpoint_translation_rmse_mm: 5.938006472330431
83
+ metric/endpoint_rotation_rmse_deg: 3.745353331000118
84
+ metric/velocity_rmse: 0.035310372233804094
85
+ metric/acceleration_rmse: 0.129455449355218
86
+ metric/jerk_rmse: 3.1107937143469813
87
+ metric/command_acceleration_rms: 0.11191823623948179
88
+ metric/command_jerk_rms: 2.6516284593632116
89
+ metric/residual_translation_mean_mm: 1.0198310279331884
90
+ metric/residual_rotation_mean_deg: 0.7992073678639711
91
+ checkpoint:
92
+ path: /home/coder/share/experiment-runs/E002/E002-R003/best.pt
93
+ sha256: 3490f23a0a5bcddb9d1998d08af37318023c8f792b17c68c1562b317dc88554a
94
+ selection_metric: loss/total
95
+ selection_split: validation
96
+ artifacts:
97
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R003/metrics.jsonl
98
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R003/evaluation_validation.json
99
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R003/condition.yaml
100
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R003/config_resolved.yaml
101
+ compute:
102
+ wall_seconds: 289.54701239199494
103
+ gpu_hours: 0.08042972566444304
104
+ peak_memory_bytes: 2951959040
experiments/E002/runs/E002-R004/condition_config.yaml ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.condition_config.v1
2
+ experiment_id: E002
3
+ run_id: E002-R004
4
+ condition_id: c-minjerk-se3-cp6-af54a01c9b
5
+ condition_summary: minjerk-se3-cp6
6
+ condition:
7
+ model:
8
+ action_head: minimum_jerk
9
+ base_conditioning: E002-R001-profiled-proposal
10
+ horizon_steps: 16
11
+ trajectory_basis: quintic_minimum_jerk_bezier_cp6
12
+ resolved_config:
13
+ data:
14
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
15
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
16
+ variant_id: validated-physical-recovery-15k-v1
17
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
18
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
19
+ sample_count: 15000
20
+ eligible_sample_count: 11628
21
+ eligible_group_count: 89
22
+ eligibility: E001 train source groups only; E001 validation and test groups excluded
23
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
24
+ split_namespace: 'E002-source-group-split-v1::'
25
+ validation_fraction: 0.1
26
+ test_fraction: 0.1
27
+ geometry:
28
+ cumulative_waypoint_representation: current_tool_relative_pose
29
+ step_representation: integrated_body_twist
30
+ step_integration: right_se3_exponential
31
+ residual_curve_frame: initial_local_tool_frame
32
+ residual_composition: base_right_multiply_se3_exponential
33
+ structured_translation: se3_tangent_translation_control_points_m
34
+ structured_rotation: so3_lie_algebra_control_points_rad
35
+ evaluation_grid: 16 positive-time waypoints at 0.05 s through 0.8 s
36
+ residual_boundary_at_t0: identity
37
+ model:
38
+ backbone:
39
+ hidden_size: 256
40
+ fusion_layers: 4
41
+ fusion_heads: 8
42
+ fusion_mlp_ratio: 4
43
+ dropout: 0.1
44
+ vision_backend: conv
45
+ text_backend: embedding
46
+ pretrained_model_name: google/siglip-base-patch16-224
47
+ freeze_pretrained: true
48
+ image_size: 224
49
+ vision_tokens: 64
50
+ depth_tokens: 36
51
+ point_tokens: 32
52
+ point_samples: 256
53
+ max_text_tokens: 32
54
+ text_vocab_size: 32768
55
+ max_joints: 7
56
+ aloha_joints: 6
57
+ panda_joints: 7
58
+ translation_scale_m: 0.05
59
+ rotation_scale_rad: 0.5235987755982988
60
+ joint_scale_rad: 0.5
61
+ horizon_steps: 16
62
+ horizon_seconds: 0.8
63
+ decoder_layers: 4
64
+ decoder_heads: 8
65
+ decoder_mlp_ratio: 4
66
+ control_points: 6
67
+ maximum_step_translation_m: 0.01
68
+ maximum_step_rotation_rad: 0.2
69
+ maximum_residual_translation_m: 0.05
70
+ maximum_residual_rotation_rad: 0.52
71
+ action_head: minimum_jerk
72
+ base_conditioning: E002-R001-profiled-proposal
73
+ trajectory_basis: quintic_minimum_jerk_bezier_cp6
74
+ loss:
75
+ path_translation: 1.0
76
+ path_rotation: 1.0
77
+ endpoint_translation: 2.0
78
+ endpoint_rotation: 2.0
79
+ velocity: 1.0
80
+ acceleration: 0.1
81
+ jerk: 0.05
82
+ residual_magnitude: 0.02
83
+ training:
84
+ seed: 42
85
+ batch_size: 64
86
+ gradient_accumulation_steps: 1
87
+ epochs: 100
88
+ max_steps: 3000
89
+ learning_rate: 0.0003
90
+ weight_decay: 0.05
91
+ warmup_fraction: 0.05
92
+ gradient_clip_norm: 1.0
93
+ precision: bf16
94
+ workers: 8
95
+ log_every_steps: 10
96
+ validation_every_steps: 250
97
+ checkpoint_every_steps: 500
98
+ evaluation:
99
+ version: se3-spline-eval-v1
100
+ trajectory_loss:
101
+ path_translation: 1.0
102
+ path_rotation: 1.0
103
+ endpoint_translation: 2.0
104
+ endpoint_rotation: 2.0
105
+ velocity: 1.0
106
+ acceleration: 0.1
107
+ jerk: 0.05
108
+ residual_magnitude: 0.02
109
+ selection_split: validation
110
+ test_firewall_opened: false
111
+ group_bootstrap_seed: 20260810
112
+ group_bootstrap_resamples: 10000
experiments/E002/runs/E002-R004/evaluation_validation.json ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_proposal_manifest": {
3
+ "path": "/home/coder/share/experiment-artifacts/E002/R001-profiled-proposals/manifest.json",
4
+ "sha256": "7fb4fb67600a0f1b599aac4f5857a6af48513a9fcef9c348be90fe264af411a3"
5
+ },
6
+ "checkpoint": {
7
+ "path": "/home/coder/share/experiment-runs/E002/E002-R004/best.pt",
8
+ "sha256": "4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9"
9
+ },
10
+ "checkpoint_run": {
11
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
12
+ "experiment_id": "E002",
13
+ "git_commit": "033cb092edc8a14430f35e5c78e06323ee14655a",
14
+ "manifest": "/home/coder/share/experiment-runs/E002/E002-R004/run_manifest.yaml",
15
+ "manifest_sha256": "846572232e745a74451e055f38cade8d2aad594999a71cf45d7cd5ba401406ac",
16
+ "run_id": "E002-R004"
17
+ },
18
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
19
+ "data_extraction_receipt": {
20
+ "path": "/home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json",
21
+ "sha256": "ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd"
22
+ },
23
+ "experiment_id": "E002",
24
+ "group_bootstrap_95ci": {
25
+ "confidence": 0.95,
26
+ "group_count": 9,
27
+ "intervals": {
28
+ "acceleration_rmse": {
29
+ "lower": 0.10308757486442725,
30
+ "upper": 0.1397645959837569
31
+ },
32
+ "command_acceleration_rms": {
33
+ "lower": 0.08936153536455499,
34
+ "upper": 0.11839342810627487
35
+ },
36
+ "command_jerk_rms": {
37
+ "lower": 1.719706772930092,
38
+ "upper": 2.6594989677270253
39
+ },
40
+ "endpoint_rotation_rmse_deg": {
41
+ "lower": 3.205244329240587,
42
+ "upper": 4.427248266008165
43
+ },
44
+ "endpoint_translation_rmse_mm": {
45
+ "lower": 5.687558265527089,
46
+ "upper": 6.8257540742556255
47
+ },
48
+ "group_macro_path_score": {
49
+ "lower": 0.7683746231926811,
50
+ "upper": 0.9666738802856869
51
+ },
52
+ "jerk_rmse": {
53
+ "lower": 2.1001228921943245,
54
+ "upper": 3.2242833753426865
55
+ },
56
+ "path_rotation_rmse_deg": {
57
+ "lower": 2.028298435277409,
58
+ "upper": 2.7444795823759502
59
+ },
60
+ "path_translation_rmse_mm": {
61
+ "lower": 3.546643653843138,
62
+ "upper": 4.226622504658169
63
+ },
64
+ "residual_rotation_mean_deg": {
65
+ "lower": 0.5349525012903743,
66
+ "upper": 0.5974235445261001
67
+ },
68
+ "residual_translation_mean_mm": {
69
+ "lower": 0.7040380966332224,
70
+ "upper": 0.9369111968411339
71
+ },
72
+ "velocity_rmse": {
73
+ "lower": 0.03159253713157442,
74
+ "upper": 0.041467026931544144
75
+ }
76
+ },
77
+ "method": "ordinary-percentile-bootstrap-over-source-groups",
78
+ "resamples": 10000,
79
+ "seed": 20260810
80
+ },
81
+ "group_count": 9,
82
+ "grouped_metrics": {
83
+ "acceleration_rmse": 0.12487276643514633,
84
+ "command_acceleration_rms": 0.10622868686914444,
85
+ "command_jerk_rms": 2.2715606689453125,
86
+ "endpoint_rotation_rmse_deg": 3.7452497482299805,
87
+ "endpoint_translation_rmse_mm": 6.209352493286133,
88
+ "jerk_rmse": 2.7772092819213867,
89
+ "path_rotation_rmse_deg": 2.348088264465332,
90
+ "path_translation_rmse_mm": 3.861748695373535,
91
+ "residual_rotation_mean_deg": 0.5649371147155762,
92
+ "residual_translation_mean_mm": 0.8120338320732117,
93
+ "velocity_rmse": 0.0359807126224041
94
+ },
95
+ "latency": {
96
+ "device": "cuda",
97
+ "median_ms": 8.1037045,
98
+ "p95_ms": 8.980220799999982,
99
+ "p99_ms": 10.40966553,
100
+ "repeats": 500,
101
+ "warmup": 50
102
+ },
103
+ "losses": {
104
+ "acceleration": 0.0013029025867581367,
105
+ "endpoint_rotation": 0.10964572429656982,
106
+ "endpoint_translation": 0.0024704034440219402,
107
+ "jerk": 0.0016885013319551945,
108
+ "path_rotation": 0.06242626905441284,
109
+ "path_translation": 0.0009601161582395434,
110
+ "residual_magnitude": 0.03481997549533844,
111
+ "total": 0.34016457200050354,
112
+ "velocity": 0.05163482204079628
113
+ },
114
+ "metrics": {
115
+ "acceleration_rmse": 0.1349928230047226,
116
+ "command_acceleration_rms": 0.11819975078105927,
117
+ "command_jerk_rms": 2.664361000061035,
118
+ "endpoint_rotation_rmse_deg": 3.799849510192871,
119
+ "endpoint_translation_rmse_mm": 6.087368965148926,
120
+ "jerk_rmse": 3.123729705810547,
121
+ "path_rotation_rmse_deg": 2.3968474864959717,
122
+ "path_translation_rmse_mm": 3.7949631214141846,
123
+ "residual_rotation_mean_deg": 0.5683494806289673,
124
+ "residual_translation_mean_mm": 0.7937496900558472,
125
+ "velocity_rmse": 0.03653155639767647
126
+ },
127
+ "per_group_metrics": {
128
+ "rlbench::rlbench/usb::datasets/rlbench_expert_validation/samples/insert_usb_in_computer/episode_000000/trajectory.npz": {
129
+ "acceleration_rmse": 0.1344926357269287,
130
+ "command_acceleration_rms": 0.12197227776050568,
131
+ "command_jerk_rms": 2.801678419113159,
132
+ "endpoint_rotation_rmse_deg": 4.230731964111328,
133
+ "endpoint_translation_rmse_mm": 6.085890293121338,
134
+ "jerk_rmse": 3.1305017471313477,
135
+ "path_rotation_rmse_deg": 2.674309253692627,
136
+ "path_translation_rmse_mm": 3.8025522232055664,
137
+ "residual_rotation_mean_deg": 0.5871337652206421,
138
+ "residual_translation_mean_mm": 0.9068980813026428,
139
+ "velocity_rmse": 0.040271393954753876
140
+ },
141
+ "robotwin_generic::002_bowl/base1::candidate_0001": {
142
+ "acceleration_rmse": 0.1561744213104248,
143
+ "command_acceleration_rms": 0.1311158388853073,
144
+ "command_jerk_rms": 3.008500099182129,
145
+ "endpoint_rotation_rmse_deg": 3.859532117843628,
146
+ "endpoint_translation_rmse_mm": 7.020136833190918,
147
+ "jerk_rmse": 3.663717269897461,
148
+ "path_rotation_rmse_deg": 2.4413540363311768,
149
+ "path_translation_rmse_mm": 4.408639907836914,
150
+ "residual_rotation_mean_deg": 0.5009645819664001,
151
+ "residual_translation_mean_mm": 0.7318453788757324,
152
+ "velocity_rmse": 0.03755943849682808
153
+ },
154
+ "robotwin_generic::050_bell/base0::candidate_0018": {
155
+ "acceleration_rmse": 0.12712980806827545,
156
+ "command_acceleration_rms": 0.11159520596265793,
157
+ "command_jerk_rms": 2.3976645469665527,
158
+ "endpoint_rotation_rmse_deg": 3.7091081142425537,
159
+ "endpoint_translation_rmse_mm": 5.565788745880127,
160
+ "jerk_rmse": 2.9391865730285645,
161
+ "path_rotation_rmse_deg": 2.3155221939086914,
162
+ "path_translation_rmse_mm": 3.4275779724121094,
163
+ "residual_rotation_mean_deg": 0.5825850367546082,
164
+ "residual_translation_mean_mm": 0.713529646396637,
165
+ "velocity_rmse": 0.03522006422281265
166
+ },
167
+ "robotwin_generic::051_candlestick/base0::candidate_0019": {
168
+ "acceleration_rmse": 0.04633811488747597,
169
+ "command_acceleration_rms": 0.046336181461811066,
170
+ "command_jerk_rms": 0.2913300395011902,
171
+ "endpoint_rotation_rmse_deg": 6.0872697830200195,
172
+ "endpoint_translation_rmse_mm": 8.130516052246094,
173
+ "jerk_rmse": 0.2913312017917633,
174
+ "path_rotation_rmse_deg": 3.6939315795898438,
175
+ "path_translation_rmse_mm": 4.995799541473389,
176
+ "residual_rotation_mean_deg": 0.5199920535087585,
177
+ "residual_translation_mean_mm": 1.134708046913147,
178
+ "velocity_rmse": 0.05466964840888977
179
+ },
180
+ "robotwin_generic::063_tabletrashbin/base10::candidate_0081": {
181
+ "acceleration_rmse": 0.13279545307159424,
182
+ "command_acceleration_rms": 0.11015186458826065,
183
+ "command_jerk_rms": 2.5362021923065186,
184
+ "endpoint_rotation_rmse_deg": 2.6013705730438232,
185
+ "endpoint_translation_rmse_mm": 6.034164905548096,
186
+ "jerk_rmse": 3.1239142417907715,
187
+ "path_rotation_rmse_deg": 1.6837923526763916,
188
+ "path_translation_rmse_mm": 3.8183398246765137,
189
+ "residual_rotation_mean_deg": 0.5236051082611084,
190
+ "residual_translation_mean_mm": 0.7148460149765015,
191
+ "velocity_rmse": 0.026725247502326965
192
+ },
193
+ "robotwin_generic::068_boxdrink/base2::candidate_0027": {
194
+ "acceleration_rmse": 0.1420787125825882,
195
+ "command_acceleration_rms": 0.09741360694169998,
196
+ "command_jerk_rms": 2.0236129760742188,
197
+ "endpoint_rotation_rmse_deg": 2.8853201866149902,
198
+ "endpoint_translation_rmse_mm": 4.953864574432373,
199
+ "jerk_rmse": 3.1655778884887695,
200
+ "path_rotation_rmse_deg": 1.8346418142318726,
201
+ "path_translation_rmse_mm": 3.1223437786102295,
202
+ "residual_rotation_mean_deg": 0.6529508829116821,
203
+ "residual_translation_mean_mm": 1.0946613550186157,
204
+ "velocity_rmse": 0.02910573035478592
205
+ },
206
+ "robotwin_generic::071_can/base1::candidate_0084": {
207
+ "acceleration_rmse": 0.12845297157764435,
208
+ "command_acceleration_rms": 0.1191152036190033,
209
+ "command_jerk_rms": 2.6901283264160156,
210
+ "endpoint_rotation_rmse_deg": 3.2490086555480957,
211
+ "endpoint_translation_rmse_mm": 6.103634834289551,
212
+ "jerk_rmse": 2.9270081520080566,
213
+ "path_rotation_rmse_deg": 2.0134360790252686,
214
+ "path_translation_rmse_mm": 3.7747769355773926,
215
+ "residual_rotation_mean_deg": 0.5559494495391846,
216
+ "residual_translation_mean_mm": 0.6914225816726685,
217
+ "velocity_rmse": 0.031804319471120834
218
+ },
219
+ "robotwin_generic::107_soap/base1::candidate_0106": {
220
+ "acceleration_rmse": 0.13067100942134857,
221
+ "command_acceleration_rms": 0.10602162033319473,
222
+ "command_jerk_rms": 2.293989419937134,
223
+ "endpoint_rotation_rmse_deg": 3.491788625717163,
224
+ "endpoint_translation_rmse_mm": 5.5043182373046875,
225
+ "jerk_rmse": 2.970918893814087,
226
+ "path_rotation_rmse_deg": 2.2204670906066895,
227
+ "path_translation_rmse_mm": 3.416995048522949,
228
+ "residual_rotation_mean_deg": 0.5371366143226624,
229
+ "residual_translation_mean_mm": 0.600003719329834,
230
+ "velocity_rmse": 0.03396544232964516
231
+ },
232
+ "robotwin_generic::111_callbell/base2::candidate_0109": {
233
+ "acceleration_rmse": 0.12572182714939117,
234
+ "command_acceleration_rms": 0.11233633756637573,
235
+ "command_jerk_rms": 2.4009382724761963,
236
+ "endpoint_rotation_rmse_deg": 3.593118906021118,
237
+ "endpoint_translation_rmse_mm": 6.4858598709106445,
238
+ "jerk_rmse": 2.7827272415161133,
239
+ "path_rotation_rmse_deg": 2.255338430404663,
240
+ "path_translation_rmse_mm": 3.9887139797210693,
241
+ "residual_rotation_mean_deg": 0.6241169571876526,
242
+ "residual_translation_mean_mm": 0.7203899025917053,
243
+ "velocity_rmse": 0.034505102783441544
244
+ }
245
+ },
246
+ "prediction_source": "checkpoint",
247
+ "primary_path_score": {
248
+ "definition": "path_translation_rmse_mm / 10 + path_rotation_rmse_deg / 5",
249
+ "group_macro": 0.8557925224304199,
250
+ "micro": 0.8588658094406127
251
+ },
252
+ "run_id": "E002-R004",
253
+ "sample_count": 1298,
254
+ "schema_version": "precision.action_chunk.evaluation.v1",
255
+ "source_family_metrics": {
256
+ "rlbench": {
257
+ "group_count": 1,
258
+ "metrics": {
259
+ "acceleration_rmse": 0.1344926357269287,
260
+ "command_acceleration_rms": 0.12197227776050568,
261
+ "command_jerk_rms": 2.801678419113159,
262
+ "endpoint_rotation_rmse_deg": 4.230731964111328,
263
+ "endpoint_translation_rmse_mm": 6.085890293121338,
264
+ "jerk_rmse": 3.1305017471313477,
265
+ "path_rotation_rmse_deg": 2.674309253692627,
266
+ "path_translation_rmse_mm": 3.8025522232055664,
267
+ "residual_rotation_mean_deg": 0.5871337652206421,
268
+ "residual_translation_mean_mm": 0.9068980813026428,
269
+ "velocity_rmse": 0.040271393954753876
270
+ },
271
+ "path_score": 0.9151170730590821,
272
+ "sample_count": 520
273
+ },
274
+ "robotwin_generic": {
275
+ "group_count": 8,
276
+ "metrics": {
277
+ "acceleration_rmse": 0.135326087474823,
278
+ "command_acceleration_rms": 0.11560964584350586,
279
+ "command_jerk_rms": 2.568490743637085,
280
+ "endpoint_rotation_rmse_deg": 3.482255220413208,
281
+ "endpoint_translation_rmse_mm": 6.0883564949035645,
282
+ "jerk_rmse": 3.119194984436035,
283
+ "path_rotation_rmse_deg": 2.191901683807373,
284
+ "path_translation_rmse_mm": 3.789881944656372,
285
+ "residual_rotation_mean_deg": 0.5557944178581238,
286
+ "residual_translation_mean_mm": 0.7181235551834106,
287
+ "velocity_rmse": 0.03380199894309044
288
+ },
289
+ "path_score": 0.8173685312271117,
290
+ "sample_count": 778
291
+ }
292
+ },
293
+ "split": "validation",
294
+ "test_firewall_opened": false
295
+ }
experiments/E002/runs/E002-R004/run_manifest.yaml ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: upvla.experiment_run.v1
2
+ run:
3
+ id: E002-R004
4
+ experiment_id: E002
5
+ condition_id: c-minjerk-se3-cp6-af54a01c9b
6
+ status: completed
7
+ started_at: '2026-08-10T11:42:37.552494+00:00'
8
+ ended_at: '2026-08-10T11:47:36.515256+00:00'
9
+ condition:
10
+ model:
11
+ action_head: minimum_jerk
12
+ base_conditioning: E002-R001-profiled-proposal
13
+ horizon_steps: 16
14
+ trajectory_basis: quintic_minimum_jerk_bezier_cp6
15
+ randomness:
16
+ global_seed: 42
17
+ python_seed: 42
18
+ numpy_seed: 42
19
+ torch_seed: 42
20
+ dataloader_seed: 42
21
+ code:
22
+ repo: https://github.com/Shiki42/UP-VLA.git
23
+ git_commit: 033cb092edc8a14430f35e5c78e06323ee14655a
24
+ git_branch: shuyuan/se3-bspline-recovery-e002
25
+ dirty: false
26
+ status_sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
27
+ patch_ref: null
28
+ data:
29
+ hf_repo_id: Shiki42/up-vla-precision-recovery-15k
30
+ revision: 77f5f306303361d602ab10759b3a139c67f2cf16
31
+ variant_id: validated-physical-recovery-15k-v1
32
+ manifest_sha256: 4cda13bc86038829f5fd519c7c766d40e770ad915974f68afea6f251b59faa50
33
+ dataset_info_sha256: 99a2f93632799a2b0a5da785438a55934e7fbaac582de0d63f8ddc0186b10ef3
34
+ extraction_receipt: /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples/_extraction_receipt.json
35
+ extraction_receipt_sha256: ff05bdb511990c0f951cb17efb1f84cc4c4a48e59f875ebd6d66a468c3b840dd
36
+ split_protocol: e001-train-only-source-family-stratified-group-hash-e002-v1
37
+ split_counts:
38
+ train_samples: 8807
39
+ validation_samples: 1298
40
+ train_groups: 71
41
+ validation_groups: 9
42
+ overlap: false
43
+ environment:
44
+ python_version: 3.11.15
45
+ python_executable: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python
46
+ packages:
47
+ numpy: 1.26.0
48
+ Pillow: 11.3.0
49
+ PyYAML: 6.0.3
50
+ torch: 2.13.0
51
+ torch_version: 2.13.0+cu130
52
+ cuda_runtime: '13.0'
53
+ cudnn_version: 92000
54
+ os: Linux-7.0.0-28-generic-x86_64-with-glibc2.35
55
+ hostname: lm
56
+ hardware:
57
+ accelerators:
58
+ - index: 0
59
+ name: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
60
+ memory_bytes: 101973491712
61
+ compute_capability: '12.0'
62
+ accelerator_count: 1
63
+ cpu_count: 32
64
+ execution:
65
+ command: /home/coder/share/experiment-runtimes/E001-py311-cu130/bin/python scripts/train_action_chunk_recovery.py
66
+ --condition-config configs/experiments/E002/R004-minjerk-se3-cp6.yaml --data-root
67
+ /home/coder/share/datasets/up-vla-precision-recovery-15k-77f5f306-samples --output-dir
68
+ /home/coder/share/experiment-runs/E002/E002-R004 --base-proposal-root /home/coder/share/experiment-artifacts/E002/R001-profiled-proposals
69
+ metrics:
70
+ validation:
71
+ loss/total: 0.340089762706786
72
+ loss/path_translation: 0.0009601259408934825
73
+ loss/path_rotation: 0.062407021208426616
74
+ loss/endpoint_translation: 0.002470431426390776
75
+ loss/endpoint_rotation: 0.10961773915816161
76
+ loss/velocity: 0.051635161111094
77
+ loss/acceleration: 0.0013029350546147037
78
+ loss/jerk: 0.0016885478681280565
79
+ loss/residual_magnitude: 0.034819636315979834
80
+ metric/path_translation_rmse_mm: 3.7914421139586323
81
+ metric/path_rotation_rmse_deg: 2.3811969316244492
82
+ metric/endpoint_translation_rmse_mm: 6.081028670118476
83
+ metric/endpoint_rotation_rmse_deg: 3.774994674559183
84
+ metric/velocity_rmse: 0.036343481690150746
85
+ metric/acceleration_rmse: 0.13480377996280124
86
+ metric/jerk_rmse: 3.113585577540111
87
+ metric/command_acceleration_rms: 0.11796897273217952
88
+ metric/command_jerk_rms: 2.653614019576133
89
+ metric/residual_translation_mean_mm: 0.7937416224891122
90
+ metric/residual_rotation_mean_deg: 0.5683441197008858
91
+ checkpoint:
92
+ path: /home/coder/share/experiment-runs/E002/E002-R004/best.pt
93
+ sha256: 4538d12581ba6bb7ecaa2d0d1a90aa7a0c265c3daafdbd7a8d5da228f9849ba9
94
+ selection_metric: loss/total
95
+ selection_split: validation
96
+ artifacts:
97
+ metrics_jsonl: /home/coder/share/experiment-runs/E002/E002-R004/metrics.jsonl
98
+ validation_evaluation: /home/coder/share/experiment-runs/E002/E002-R004/evaluation_validation.json
99
+ condition_snapshot: /home/coder/share/experiment-runs/E002/E002-R004/condition.yaml
100
+ resolved_config: /home/coder/share/experiment-runs/E002/E002-R004/config_resolved.yaml
101
+ compute:
102
+ wall_seconds: 298.82912589100306
103
+ gpu_hours: 0.08300809052527863
104
+ peak_memory_bytes: 2951936512
experiments/E002/selection-validation.json ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "candidate_gates": {
3
+ "c-cubic-se3-cp6-43b928e718": {
4
+ "batch1_gpu_p95_below_20ms": true,
5
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
6
+ "endpoint_translation_within_endpoint_plus_1mm": true,
7
+ "jerk_not_above_endpoint_by_10pct": true,
8
+ "jerk_reduced_at_least_25pct_vs_residual": true,
9
+ "paired_jerk_ratio_95ci": {
10
+ "lower": 0.6329302244799777,
11
+ "method": "paired-percentile-bootstrap-over-source-groups",
12
+ "point": 0.7373313062669027,
13
+ "resamples": 10000,
14
+ "seed": 20260810,
15
+ "upper": 0.8095807915029037
16
+ },
17
+ "paired_jerk_ratio_95ci_upper_below_1": true,
18
+ "path_not_worse_than_residual_by_2pct": true
19
+ },
20
+ "c-minjerk-se3-cp6-af54a01c9b": {
21
+ "batch1_gpu_p95_below_20ms": true,
22
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
23
+ "endpoint_translation_within_endpoint_plus_1mm": true,
24
+ "jerk_not_above_endpoint_by_10pct": true,
25
+ "jerk_reduced_at_least_25pct_vs_residual": true,
26
+ "paired_jerk_ratio_95ci": {
27
+ "lower": 0.5945370613267315,
28
+ "method": "paired-percentile-bootstrap-over-source-groups",
29
+ "point": 0.7240625802894608,
30
+ "resamples": 10000,
31
+ "seed": 20260810,
32
+ "upper": 0.8080530623282974
33
+ },
34
+ "paired_jerk_ratio_95ci_upper_below_1": true,
35
+ "path_not_worse_than_residual_by_2pct": true
36
+ }
37
+ },
38
+ "endpoint_condition_id": "c-endpoint-5d2dff166d",
39
+ "experiment_id": "E002",
40
+ "gates": {
41
+ "batch1_gpu_p95_below_20ms": true,
42
+ "endpoint_rotation_within_endpoint_plus_1deg": true,
43
+ "endpoint_translation_within_endpoint_plus_1mm": true,
44
+ "jerk_not_above_endpoint_by_10pct": true,
45
+ "jerk_reduced_at_least_25pct_vs_residual": true,
46
+ "paired_jerk_ratio_95ci": {
47
+ "lower": 0.5945370613267315,
48
+ "method": "paired-percentile-bootstrap-over-source-groups",
49
+ "point": 0.7240625802894608,
50
+ "resamples": 10000,
51
+ "seed": 20260810,
52
+ "upper": 0.8080530623282974
53
+ },
54
+ "paired_jerk_ratio_95ci_upper_below_1": true,
55
+ "path_not_worse_than_residual_by_2pct": true
56
+ },
57
+ "inputs": {
58
+ "candidates": [
59
+ {
60
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R003-validation.json",
61
+ "sha256": "237728c0a5594821ef420a5ec4452943767845a7058ccf32f7dc6515ab7cd330"
62
+ },
63
+ {
64
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R004-validation.json",
65
+ "sha256": "7c451bf90cb7e2f1ea50a3dff1647bca8a9cea6d4adf9003733d8140e6beb741"
66
+ }
67
+ ],
68
+ "endpoint": {
69
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R001-validation.json",
70
+ "sha256": "0fda3f48a9f3ac61cdc91d4e13c0cd0d7bbb006c23a4f074d0426313da506b06"
71
+ },
72
+ "experiment": {
73
+ "path": "/home/coder/share/UP-VLA-se3-bspline-E002/experiments/E002-se3-bspline-minjerk-recovery.yaml",
74
+ "sha256": "fc288032290f2706bd04f58004d5df7166f7823695ad715e44e0bb1bc5bb5417"
75
+ },
76
+ "residual": {
77
+ "path": "/home/coder/share/experiment-artifacts/E002/evaluations/R002-validation.json",
78
+ "sha256": "17a116944f1967d60d2427975690f635f4886180578b92b969be7d9d756c84a1"
79
+ }
80
+ },
81
+ "ranking": [
82
+ {
83
+ "batch1_gpu_p95_ms": 8.980220799999982,
84
+ "command_jerk_rms": 2.2715606689453125,
85
+ "condition_id": "c-minjerk-se3-cp6-af54a01c9b",
86
+ "endpoint_rotation_rmse_deg": 3.7452497482299805,
87
+ "endpoint_translation_rmse_mm": 6.209352493286133,
88
+ "group_macro_path_score": 0.8557925224304199,
89
+ "run_id": "E002-R004"
90
+ },
91
+ {
92
+ "batch1_gpu_p95_ms": 8.21552085,
93
+ "command_jerk_rms": 2.313187599182129,
94
+ "condition_id": "c-cubic-se3-cp6-43b928e718",
95
+ "endpoint_rotation_rmse_deg": 3.8112881183624268,
96
+ "endpoint_translation_rmse_mm": 6.002523422241211,
97
+ "group_macro_path_score": 0.8477689743041992,
98
+ "run_id": "E002-R003"
99
+ }
100
+ ],
101
+ "residual_condition_id": "c-residual-h16-f186e90df2",
102
+ "schema_version": "precision.e002.selection.v1",
103
+ "selected_condition_id": "c-minjerk-se3-cp6-af54a01c9b",
104
+ "selected_run_id": "E002-R004",
105
+ "selection_rule": "minimum grouped command jerk among path/endpoint/latency-guarded candidates; if none are guarded, minimum grouped command jerk among all candidates",
106
+ "selection_split": "validation",
107
+ "test_firewall_opened": false,
108
+ "validation_gate_passed": true
109
+ }