Promote robust iteration 12195 running policy

.gitattributes

@@ -1,3 +1,6 @@
 *.onnx filter=lfs diff=lfs merge=lfs -text
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 
+checkpoint.pt filter=lfs diff=lfs merge=lfs -text
+legacy/iteration-8749/checkpoint.pt filter=lfs diff=lfs merge=lfs -text
+lineage/iteration-11748/checkpoint.pt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -6,43 +6,53 @@ license: apache-2.0
 
 # microduck-running
 
-Run forward on flat ground with a stock MicroDuck. This is the iteration-8,749
-policy used in the run-into-mat video, not the later faster continuation.
+A robustified forward-running policy for the stock MicroDuck on flat ground.
+The default artifact is iteration **12,195**, continued from the 1.687 m/s
+speed frontier with progressively stronger velocity pushes, centre-of-mass
+variation, and initial tilt.
 
-**Status: simulation only; not yet validated on hardware.** It is a sim-to-real
-candidate, not a hardware-ready claim. The policy reaches a mean 1.61 m/s
-body-forward speed at the recommended 1.80 m/s simulation command, but its
-heading control is still weak.
+**Status: simulation-only hardware candidate; not validated on hardware.** It
+retains 1.651 m/s mean body-forward speed in the nominal simulation battery and
+1.612 m/s under the combined backlash-plus-disturbance battery. Heading and
+lateral drift remain substantial.
 
-## Normal simulation rollout
+## Normal simulation rollout — iteration 12,195
 
 <video src="https://huggingface.co/HannesVonEssen/microduck-running/resolve/main/media/preview.mp4" width="720" controls autoplay loop muted playsinline preload="metadata"></video>
 
 [Open the full silent MP4](media/preview.mp4).
 
-## Run-into-mat edit
+## Run-into-mat edit — legacy iteration 8,749
 
 <video src="https://huggingface.co/HannesVonEssen/microduck-running/resolve/main/media/run-into-mat.mp4" controls loop muted playsinline></video>
 
 [Open the full silent run-into-mat MP4](media/run-into-mat.mp4).
 
-The second video uses the same running checkpoint, but the mat, full-body
-collision additions, fixed-camera treatment, post-impact motor cutoff, and
-sparks are scripted demonstration elements. The ONNX policy has no obstacle
-perception and did not learn to run into, react to, or stop at a mat. Both
-videos are silent.
+The mat video uses the preserved iteration-8,749 controller, not the new root
+policy. Its obstacle, collision augmentation, camera treatment, post-impact
+motor cutoff, and sparks are scripted demonstration elements. The policy has
+no obstacle perception and did not learn to react to a mat. Both videos are
+silent.
+
+## Versions
+
+| Role | Iteration | Files | Note |
+|---|---:|---|---|
+| Default robust candidate | 12,195 | [`policy.onnx`](policy.onnx), [`checkpoint.pt`](checkpoint.pt) | Recommended starting point for further simulation and cautious hardware research |
+| Speed-focused parent | 11,748 | [`lineage/iteration-11748/`](lineage/iteration-11748/) | 1.687 m/s frontier used to start robustification |
+| Legacy mat-video policy | 8,749 | [`legacy/iteration-8749/`](legacy/iteration-8749/) | Exact controller used in the run-into-mat edit |
 
 ## Command
 
 The 13-D command block is `[twist(3), head_pose(4), body_pose(6)]`.
 
-| slot | meaning |
+| Slot | Meaning |
 |---|---|
-| `twist[0]` | forward-velocity command; curriculum trained through 1.65 m/s, evaluated at 1.80 m/s |
-| `twist[1]` | lateral command; use 0 (only a tiny ±0.02 m/s range was seen in training) |
-| `twist[2]` | yaw-rate command; use 0 (only a tiny ±0.05 rad/s range was seen in training) |
-| `head_pose[0:4]` | unused for this policy; zeros |
-| `body_pose[0:6]` | unused for this policy; zeros |
+| `twist[0]` | forward-velocity command; trained and evaluated through 2.20 m/s |
+| `twist[1]` | lateral command; use 0 (only ±0.02 m/s was seen in training) |
+| `twist[2]` | yaw-rate command; use 0 (only ±0.05 rad/s was seen in training) |
+| `head_pose[0:4]` | unused; zeros |
+| `body_pose[0:6]` | unused; zeros |
 
 An exact-zero standing bucket made up 3% of training environments, so
 `twist = [0, 0, 0]` is the intended idle command. The speed objective rewards
@@ -51,35 +61,43 @@ limiter.
 
 ## Try it in simulation
 
-The matching training and evaluation source is in
-[`Vottivott/microduck-playground`](https://github.com/Vottivott/microduck-playground)
-at commit
-[`c5fcc50`](https://github.com/Vottivott/microduck-playground/commit/c5fcc50219fef01ac9931d0079c583ccbb29b689):
+The matching source is
+[`Vottivott/microduck-playground@828d950`](https://github.com/Vottivott/microduck-playground/commit/828d950134e29a8d04cbb51720a22c8729047fb7):
 
 ```bash
 git clone https://github.com/Vottivott/microduck-playground.git
 cd microduck-playground
-git checkout c5fcc50219fef01ac9931d0079c583ccbb29b689
+git checkout 828d950134e29a8d04cbb51720a22c8729047fb7
 uv sync
 
 hf download HannesVonEssen/microduck-running policy.onnx config.json --local-dir policies/running
 uv run scripts/infer_policy.py \
   --walking policies/running/policy.onnx \
   --new-cmd-obs \
-  --lin-vel-x 1.8
+  --lin-vel-x 2.2
 ```
 
 ## Continue training
 
-`checkpoint.pt` is the complete iteration-8,749 PPO state used for
-this release: actor, critic, optimizer, observation normalizers, and curriculum
-counter. Place it at
-`logs/rsl_rl/running/release-8749/model_8749.pt`, then resume with:
+`checkpoint.pt` is the complete iteration-12,195 PPO state: actor, critic,
+optimizer, observation normalizers, and curriculum counter.
 
 ```bash
+mkdir -p logs/rsl_rl/running/release-12195
+cp checkpoint.pt logs/rsl_rl/running/release-12195/model_12195.pt
+
+MICRODUCK_RUNNING_TARGET_MAX_SPEED=2.2 \
+MICRODUCK_RUNNING_SPEED_CAP=2.4 \
+MICRODUCK_RUNNING_HIGH_SPEED_STAGE_INTERVAL=750 \
+MICRODUCK_RUNNING_ACTION_RATE_WEIGHT=-0.10 \
+MICRODUCK_RUNNING_FORWARD_PROGRESS_WEIGHT=5.0 \
+MICRODUCK_RUNNING_ROBUST_PUSH_MPS=0.10 \
+MICRODUCK_RUNNING_ROBUST_TRUNK_COM_M=0.008 \
+MICRODUCK_RUNNING_ROBUST_HEAD_COM_M=0.006 \
+MICRODUCK_RUNNING_ROBUST_INITIAL_TILT_DEG=2.0 \
 uv run train Mjlab-Running-Flat-MicroDuck \
-  --agent.resume True --agent.load-run release-8749 \
-  --agent.load-checkpoint model_8749.pt --agent.max-iterations 100
+  --agent.resume True --agent.load-run release-12195 \
+  --agent.load-checkpoint model_12195.pt --agent.max-iterations 100
 ```
 
 PyTorch checkpoints use pickle internally; load them only from a repository
@@ -92,83 +110,84 @@ and revision you trust.
 - control rate: 50 Hz
 - action scale: 1.0, joint-position targets around MicroDuck HOME
 - observation normalizer: baked into `policy.onnx`
+- output clipping: none; training also used unclipped actor outputs
 - entry pose: standing
 - robot: MicroDuck hardware revision 1, 14 Dynamixel XL330 servos
-- policy type: perpetual flat-ground locomotion
+- hardware modifications: none
+
+Checkpoint-to-ONNX parity over zero and deterministic random observations had
+a maximum absolute action error of 7.63e-6. See
+[`eval/checkpoint-onnx-parity.json`](eval/checkpoint-onnx-parity.json).
 
 ## Simulation evaluation
 
-At a 1.80 m/s forward command over 256 randomized environments for 8 seconds:
+All batteries used a 2.20 m/s command, 1-second warm-up, and 10-second
+measurement horizon.
+
+| Evaluation | Environments | Body-forward speed | Survival |
+|---|---:|---:|---:|
+| Ordinary model, nominal | 512 | 1.651 m/s | 99.22% |
+| Ordinary model, push/CoM/tilt stress | 512 | 1.635 m/s | 98.83% |
+| Backlash model, nominal | 256 | 1.636 m/s | 98.83% |
+| Backlash model, push/CoM/tilt stress | 256 | 1.612 m/s | 98.44% |
+| High grip plus full stress | 512 | 1.642 m/s | 96.68% |
 
-| metric | result |
-|---|---:|
-| survival | 99.22% |
-| mean body-forward speed | 1.6073 m/s |
-| mean displacement speed along initial heading | 1.3277 m/s |
-| mean absolute heading error | 25.22° |
-| mean absolute lateral displacement | 3.87 m |
-| flight fraction | 65.72% |
+Stress means ±0.10 m/s planar pushes every 3–6 seconds, ±10 mm trunk CoM,
+±6 mm head CoM, and ±2° initial pitch/roll. The high-grip case widens foot
+friction from 0.7–1.3 to 0.7–1.8. Raw outputs are under [`eval/`](eval/).
 
-At a 2.00 m/s extrapolated command, mean body-forward speed rose only to
-1.6452 m/s while survival fell to 93.36%. Do not treat a higher command as a
-safe route to higher speed. The `1.6 m/s` video label is the measured
-simulation body-forward mean at the 1.80 m/s command; it is not a real-hardware
-measurement. `WALKING BASELINE 0.4 m/s` refers to the walking task's configured
-command ceiling, not a matched measured baseline.
+The candidate was trained with the ordinary robot model, then evaluated
+unchanged with the backlash model. A separately backlash-trained continuation
+performed worse and was rejected. Robustification costs about 2.1% nominal
+speed relative to iteration 11,748 but gives the stronger overall release.
 
-## Sim-to-real constraints
+## Sim-to-real boundary
 
-The run kept important deployment-oriented modeling from the MicroDuck walking
-recipe:
+The policy retains substantial deployment-oriented modeling:
 
-- BAM XL330 M6 voltage control, back-EMF, current limiting, battery variation
-  from 6.5–8.2 V, and load-dependent voltage sag;
+- BAM XL330 M6 voltage control, back-EMF, current limiting, 6.5–8.2 V battery
+  variation, and load-dependent sag;
 - 3–6 physics-step actuator delay, 0–1-step IMU delay, and one-control-step
   joint-velocity lag;
 - foot friction randomized from 0.7–1.3;
-- trunk mass/inertia ±5%, trunk and head CoM ±3 mm, armature ±10%, and BAM
-  joint friction ±10%;
-- joint encoder bias ±0.015 rad, IMU mounting misalignment up to 6°, and actor
-  observation noise.
-
-To discover a fast gait, it deliberately relaxed several constraints relative
-to the general walking recipe:
-
-- flat terrain only and no random push disturbances;
-- the non-backlash robot model;
-- no curricula widening trunk/head CoM or head/body pose-command ranges;
-- weaker upright, pose, angular-motion, foot-slip, and action-rate
-  regularization; head/body pose tracking disabled;
-- motor-gain randomization and randomized initial base tilt were already
-  disabled in the inherited base recipe.
-
-This combination makes transfer plausible, not proven. The main known concerns
-are impact loading, unmodeled gearbox backlash/compliance, real foot contact,
-heading drift, and the relaxed robustness training.
+- trunk mass/inertia ±5%, armature ±10%, BAM joint friction ±10%, encoder bias
+  ±0.015 rad, IMU misalignment up to 6°, and actor-observation noise;
+- in the final stage: ±0.10 m/s pushes, ±8 mm trunk CoM, ±6 mm head CoM, and
+  ±2° initial pitch/roll.
+
+Important remaining relaxations are flat terrain, ordinary non-backlash
+training, disabled head/body pose tracking, weaker posture and angular-motion
+regularization, and no motor-gain randomization. Backlash was tested, not part
+of the selected policy's training distribution. Transfer is plausible, not
+proven; heading drift, impact loading, real contact, servo thermal limits, and
+unmodeled compliance remain concerns.
 
 ## Hardware safety
 
 For a first robot test, use a support/catch rig, keep a remote torque-off path,
 start with low commands and ramp conservatively, and log servo current,
 temperature, voltage, fall state, and commanded/applied twist. Do not begin at
-1.80 m/s. Stop on repeated saturation, overheating, unexpected oscillation, or
+2.20 m/s. Stop on repeated saturation, overheating, unexpected oscillation, or
 systematic forward pitching.
 
 ## Provenance and files
 
-Training task `Mjlab-Running-Flat-MicroDuck`; PPO with `rsl_rl`; run
-`sweep-speed-r1`; checkpoint iteration 8,749. Source:
-[`Vottivott/microduck-playground@c5fcc50`](https://github.com/Vottivott/microduck-playground/commit/c5fcc50219fef01ac9931d0079c583ccbb29b689),
+Training task `Mjlab-Running-Flat-MicroDuck`; PPO with `rsl_rl`; checkpoint
+iteration 12,195. Source:
+[`Vottivott/microduck-playground@828d950`](https://github.com/Vottivott/microduck-playground/commit/828d950134e29a8d04cbb51720a22c8729047fb7),
 based on
 [`pollen-robotics/microduck_rl@d424a0c`](https://github.com/pollen-robotics/microduck_rl/commit/d424a0c899f6b33cbd3daeb279913134349c0b63).
 Exported with `scripts/export.py`, which bakes the observation normalizer into
 the ONNX graph.
 
-- `policy.onnx` — deployment policy
-- `checkpoint.pt` — resumable PPO checkpoint
+- `policy.onnx` — default robust deployment policy
+- `checkpoint.pt` — default resumable PPO checkpoint
 - `manifest.json` — machine-readable contract, evaluation, and provenance
-- `media/preview.mp4` — normal 10-second close rollout
-- `media/run-into-mat.mp4` — scripted 8-second crash-mat demonstration
+- `eval/` — raw evaluation records and checkpoint/ONNX parity record
+- `media/preview.mp4` — robust policy's 10-second close rollout
+- `media/run-into-mat.mp4` — scripted legacy iteration-8,749 demonstration
+- `lineage/iteration-11748/` — speed-parent model, checkpoint, video, and evals
+- `legacy/iteration-8749/` — exact mat-video policy, checkpoint, and clean video
 - `SHA256SUMS` — artifact integrity hashes
 
 Format inspired by
@@ -176,7 +195,7 @@ Format inspired by
 
 <a href="https://hfviewer.com/HannesVonEssen/microduck-running?utm_source=huggingface&amp;utm_medium=embedded_model_card&amp;utm_campaign=HannesVonEssen__microduck-running_card&amp;utm_content=embedded_card_open_viewer&amp;from=embedded-model-card" target="_blank" rel="noopener">
   <img
-    src="https://hfviewer.com/api/card.svg?source=HannesVonEssen%2Fmicroduck-running&amp;granularity=auto&amp;v=20260902-action-contract-r2"
+    src="https://hfviewer.com/api/card.svg?source=HannesVonEssen%2Fmicroduck-running&amp;granularity=auto&amp;v=20260903-robust12195"
     alt="Architecture graph for HannesVonEssen/microduck-running. Open in hfviewer"
     width="100%"
   />

SHA256SUMS

@@ -1,7 +1,23 @@
-66a9b4ca82dbfe1376989b1f25ab382c276b31dfe96631a86447afc8e65e74ef  README.md
-ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f  checkpoint.pt
-d8923f68f7275cacaf2c436e8b222af3f9ea9b98485a4d4387865a0daabac71d  config.json
-1907518687eaccf35ef242da3326873b4c7849e31dd0dc0a677f747fc5b6a8f3  manifest.json
-69ce8b5f8855ac76d2413d90b176e770d22d22a72766dcc2dacf34ccbded58d6  media/preview.mp4
+69c602a90098db9642a213367371153a8451e0b66e21eb48d682853f411e5d97  README.md
+052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973  checkpoint.pt
+a9a1e9058642980c5f33b6c9d1ad54f513292429dc3dc57dcea8d3f6f2dc5950  config.json
+d7a92e6ed3b48e72a07973ada6b74f3611bba38aaa8af9747ba0dabaa6ef12e0  eval/checkpoint-onnx-parity.json
+f65329ffd1cfe990c244390ae076532537daee530832cf993f84d95b53e30c78  eval/high-grip-plain.json
+eafe79f0f32f0430efb13bc95bb5896d8aaf1bff575a5dc973ba30aaa2f9d162  eval/nominal-backlash.json
+ef08a1b9c1d4ac51e50d119197d2406d3352b31d7e986190449f7be058ac7236  eval/nominal-plain.json
+16918366d4208f433eabd3c24a1e57ae8923ab9fb064e70328347e64b5fd65b7  eval/released_12195.json
+a9934320b9f4bf29ff4304ec48be01c519540188cfb679b945a13721de83cb24  eval/stress-backlash.json
+abdb838ea7c4795451e97ddb27870404189f53fb1032084fa517a24e674df098  eval/stress-plain.json
+ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f  legacy/iteration-8749/checkpoint.pt
+1907518687eaccf35ef242da3326873b4c7849e31dd0dc0a677f747fc5b6a8f3  legacy/iteration-8749/manifest.json
+69ce8b5f8855ac76d2413d90b176e770d22d22a72766dcc2dacf34ccbded58d6  legacy/iteration-8749/media/preview.mp4
+b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a  legacy/iteration-8749/policy.onnx
+1e79df200d8bb19b139ee1e74e6910c62434d57e792f632b66b10a73266894eb  lineage/iteration-11748/checkpoint.pt
+a81a468c789b890f935e14f478f992c632b01eeb25179da7d3ba252d663625be  lineage/iteration-11748/evaluations.tar.gz
+04eccc1fe13a8d00c813cac78e7ada8df65ab3b317757a0b2d0a4ac35574ca8f  lineage/iteration-11748/manifest.json
+ccd6757b53b591081fdf75dd3effe9a5743f5fac0c2a513fa48dfd5a32231b65  lineage/iteration-11748/policy.onnx
+256ad675984d544204076986a7e09ea3a66c6070ec15ffb52921212e685e898c  lineage/iteration-11748/preview.mp4
+0ef9f79b9105ed28c8b4d718dd669541de6dc906493732c5357e87a8cc92a0a3  manifest.json
+942f0802cc0c6795769bd99a9991f866fee682890ea9ea6d71077f7d4eb5492b  media/preview.mp4
 13f3eaf36601928125a001c1e5f444285a7294b70dc53a186da195d8cb50bc9a  media/run-into-mat.mp4
-b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a  policy.onnx
+007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb  policy.onnx

checkpoint.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f
-size 4843445
+oid sha256:052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973
+size 4843527

config.json

@@ -8,6 +8,7 @@
   "observation_size": 61,
   "action_size": 14,
   "control_frequency_hz": 50,
+  "checkpoint_iteration": 12195,
   "source_repository": "Vottivott/microduck-playground",
-  "source_revision": "c5fcc50219fef01ac9931d0079c583ccbb29b689"
+  "source_revision": "828d950134e29a8d04cbb51720a22c8729047fb7"
 }

eval/checkpoint-onnx-parity.json

@@ -0,0 +1,13 @@
+{
+  "checkpoint_iteration": 12195,
+  "checkpoint_sha256": "052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973",
+  "onnx_sha256": "007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb",
+  "input_shape": [1, 61],
+  "output_shape": [1, 14],
+  "samples": "zero plus deterministic random observations",
+  "max_absolute_error": 0.00000762939453125,
+  "mean_absolute_error": 0.000000862,
+  "observation_normalizer_embedded": true,
+  "training_action_clipping": false,
+  "onnx_output_clip": false
+}

eval/high-grip-plain.json

@@ -0,0 +1,37 @@
+{
+  "absolute_lateral_displacement_m": {
+    "mean": 6.934412002563477,
+    "median": 7.661316394805908,
+    "p90": 11.148153305053711
+  },
+  "actor_observation_dim": 61,
+  "checkpoint": "/net/gcp/cloud-storage/europe-west/training/hannes/microduck_running/20260830-hardware-robust/v2-mild-r2/stage-c/checkpoint.pt",
+  "command_speed_mps": 2.2,
+  "duration_s": 10.0,
+  "final_absolute_heading_error_deg": {
+    "mean": 75.98717957309084,
+    "median": 70.41698272002702,
+    "p90": 149.06419312575767
+  },
+  "flight_events_per_second": 6.993881702423096,
+  "flight_fraction": 0.6284358501434326,
+  "forward_speed_mps": {
+    "clean_survivor_mean": 1.6467944383621216,
+    "mean": 1.6424154043197632,
+    "median": 1.6616889238357544,
+    "p10": 1.4980621337890625,
+    "p90": 1.7803144454956055
+  },
+  "initial_heading_displacement_speed_mps": {
+    "mean": 1.0247559547424316,
+    "median": 1.148021936416626,
+    "p10": 0.19742903113365173
+  },
+  "max_tilt_deg_median": 16.086403476606087,
+  "mean_absolute_heading_error_deg": 42.492167524874674,
+  "mean_absolute_lateral_speed_mps": 0.10666095465421677,
+  "mean_absolute_yaw_rate_radps": 1.3615901470184326,
+  "num_envs": 512,
+  "survival_fraction": 0.966796875,
+  "task_id": "Mjlab-Running-Flat-MicroDuck"
+}

eval/nominal-backlash.json

@@ -0,0 +1,37 @@
+{
+  "absolute_lateral_displacement_m": {
+    "mean": 6.03297233581543,
+    "median": 6.301413536071777,
+    "p90": 10.287883758544922
+  },
+  "actor_observation_dim": 61,
+  "checkpoint": "/net/gcp/cloud-storage/europe-west/training/hannes/microduck_running/20260830-hardware-robust/v2-mild-r2/stage-c/checkpoint.pt",
+  "command_speed_mps": 2.2,
+  "duration_s": 10.0,
+  "final_absolute_heading_error_deg": {
+    "mean": 63.66170706483097,
+    "median": 57.51983703861542,
+    "p90": 133.31073063533844
+  },
+  "flight_events_per_second": 7.0286455154418945,
+  "flight_fraction": 0.6461542844772339,
+  "forward_speed_mps": {
+    "clean_survivor_mean": 1.6365820169448853,
+    "mean": 1.6362495422363281,
+    "median": 1.64719820022583,
+    "p10": 1.515997052192688,
+    "p90": 1.7456064224243164
+  },
+  "initial_heading_displacement_speed_mps": {
+    "mean": 1.1968926191329956,
+    "median": 1.324066162109375,
+    "p10": 0.6039830446243286
+  },
+  "max_tilt_deg_median": 15.47616364030588,
+  "mean_absolute_heading_error_deg": 32.83215926473935,
+  "mean_absolute_lateral_speed_mps": 0.09709258377552032,
+  "mean_absolute_yaw_rate_radps": 1.5131072998046875,
+  "num_envs": 256,
+  "survival_fraction": 0.98828125,
+  "task_id": "Mjlab-Running-Flat-Backlash-MicroDuck"
+}

eval/nominal-plain.json

@@ -0,0 +1,37 @@
+{
+  "absolute_lateral_displacement_m": {
+    "mean": 5.830930709838867,
+    "median": 6.121419906616211,
+    "p90": 10.318323135375977
+  },
+  "actor_observation_dim": 61,
+  "checkpoint": "/net/gcp/cloud-storage/europe-west/training/hannes/microduck_running/20260830-hardware-robust/v2-mild-r2/stage-c/checkpoint.pt",
+  "command_speed_mps": 2.2,
+  "duration_s": 10.0,
+  "final_absolute_heading_error_deg": {
+    "mean": 59.87216784914615,
+    "median": 52.358908535713354,
+    "p90": 119.16307293954588
+  },
+  "flight_events_per_second": 7.0729241371154785,
+  "flight_fraction": 0.6566617488861084,
+  "forward_speed_mps": {
+    "clean_survivor_mean": 1.6519087553024292,
+    "mean": 1.6510932445526123,
+    "median": 1.6662672758102417,
+    "p10": 1.5096769332885742,
+    "p90": 1.760268211364746
+  },
+  "initial_heading_displacement_speed_mps": {
+    "mean": 1.2437968254089355,
+    "median": 1.3523855209350586,
+    "p10": 0.7378867864608765
+  },
+  "max_tilt_deg_median": 15.38079883156687,
+  "mean_absolute_heading_error_deg": 31.326498703723402,
+  "mean_absolute_lateral_speed_mps": 0.09913850575685501,
+  "mean_absolute_yaw_rate_radps": 1.3830451965332031,
+  "num_envs": 512,
+  "survival_fraction": 0.9921875,
+  "task_id": "Mjlab-Running-Flat-MicroDuck"
+}

eval/released_12195.json

@@ -0,0 +1,129 @@
+{
+  "schema_version": 1,
+  "experiment": "running",
+  "policy": {
+    "huggingface_repo": "HannesVonEssen/microduck-running",
+    "checkpoint_iteration": 12195,
+    "resumed_from_iteration": 11748,
+    "policy_onnx_sha256": "007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb",
+    "checkpoint_pt_sha256": "052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973"
+  },
+  "training": {
+    "task_id": "Mjlab-Running-Flat-MicroDuck",
+    "simulator": "mjlab / MuJoCo Warp",
+    "algorithm": "PPO (rsl_rl)",
+    "seed": 42,
+    "num_envs": 4096,
+    "common": {
+      "target_max_speed_m_s": 2.2,
+      "speed_reward_cap_m_s": 2.4,
+      "high_speed_stage_interval_iterations": 750,
+      "action_rate_reward_weight": -0.1,
+      "forward_progress_reward_weight": 5.0,
+      "symmetry_enabled": false,
+      "heading_feedback_enabled": false,
+      "push_interval_s": [3.0, 6.0],
+      "foot_friction_range": [0.7, 1.3]
+    },
+    "stages": [
+      {
+        "iterations": 100,
+        "push_velocity_m_s": 0.03,
+        "trunk_com_offset_m": 0.003,
+        "head_com_offset_m": 0.003,
+        "initial_pitch_roll_deg": 1.0
+      },
+      {
+        "iterations": 150,
+        "push_velocity_m_s": 0.06,
+        "trunk_com_offset_m": 0.005,
+        "head_com_offset_m": 0.005,
+        "initial_pitch_roll_deg": 1.5
+      },
+      {
+        "iterations": 200,
+        "push_velocity_m_s": 0.1,
+        "trunk_com_offset_m": 0.008,
+        "head_com_offset_m": 0.006,
+        "initial_pitch_roll_deg": 2.0
+      }
+    ]
+  },
+  "evaluation": {
+    "command_forward_m_s": 2.2,
+    "duration_s": 10.0,
+    "warmup_s": 1.0,
+    "cases": [
+      {
+        "name": "nominal",
+        "task_id": "Mjlab-Running-Flat-MicroDuck",
+        "num_envs": 512,
+        "survival_fraction": 0.9921875,
+        "mean_body_forward_speed_m_s": 1.6510932445526123,
+        "mean_initial_heading_displacement_speed_m_s": 1.2437968254089355,
+        "mean_absolute_heading_error_deg": 31.326498703723402
+      },
+      {
+        "name": "push-com-tilt-stress",
+        "task_id": "Mjlab-Running-Flat-MicroDuck",
+        "num_envs": 512,
+        "survival_fraction": 0.98828125,
+        "mean_body_forward_speed_m_s": 1.6349653005599976,
+        "mean_initial_heading_displacement_speed_m_s": 1.0107059478759766,
+        "mean_absolute_heading_error_deg": 43.2924983559655
+      },
+      {
+        "name": "backlash-nominal",
+        "task_id": "Mjlab-Running-Flat-Backlash-MicroDuck",
+        "num_envs": 256,
+        "survival_fraction": 0.98828125,
+        "mean_body_forward_speed_m_s": 1.6362495422363281,
+        "mean_initial_heading_displacement_speed_m_s": 1.1968926191329956,
+        "mean_absolute_heading_error_deg": 32.83215926473935
+      },
+      {
+        "name": "backlash-push-com-tilt-stress",
+        "task_id": "Mjlab-Running-Flat-Backlash-MicroDuck",
+        "num_envs": 256,
+        "survival_fraction": 0.984375,
+        "mean_body_forward_speed_m_s": 1.6121970415115356,
+        "mean_initial_heading_displacement_speed_m_s": 0.9829199910163879,
+        "mean_absolute_heading_error_deg": 43.80756975146533
+      },
+      {
+        "name": "high-grip-push-com-tilt-stress",
+        "task_id": "Mjlab-Running-Flat-MicroDuck",
+        "num_envs": 512,
+        "survival_fraction": 0.966796875,
+        "mean_body_forward_speed_m_s": 1.6424154043197632,
+        "mean_initial_heading_displacement_speed_m_s": 1.0247559547424316,
+        "mean_absolute_heading_error_deg": 42.492167524874674,
+        "foot_friction_range": [0.7, 1.8]
+      }
+    ],
+    "stress_definition": {
+      "push_velocity_m_s": 0.1,
+      "trunk_com_offset_m": 0.01,
+      "head_com_offset_m": 0.006,
+      "initial_pitch_roll_deg": 2.0
+    }
+  },
+  "checkpoint_onnx_parity": {
+    "observation_shape": [1, 61],
+    "action_shape": [1, 14],
+    "samples": "zero plus deterministic random observations",
+    "maximum_absolute_error": 0.00000762939453125,
+    "mean_absolute_error": 0.000000862,
+    "action_clipping_during_training": false,
+    "observation_normalizer_embedded": true
+  },
+  "media": {
+    "path": "../media/preview.mp4",
+    "sha256": "942f0802cc0c6795769bd99a9991f866fee682890ea9ea6d71077f7d4eb5492b",
+    "duration_s": 10.0,
+    "fps": 50,
+    "audio": false
+  },
+  "known_limit": "Heading and lateral drift remain substantial; this has not been validated on hardware.",
+  "status": "simulation-only-hardware-candidate"
+}

eval/stress-backlash.json

@@ -0,0 +1,37 @@
+{
+  "absolute_lateral_displacement_m": {
+    "mean": 6.875430107116699,
+    "median": 7.540812969207764,
+    "p90": 11.18838882446289
+  },
+  "actor_observation_dim": 61,
+  "checkpoint": "/net/gcp/cloud-storage/europe-west/training/hannes/microduck_running/20260830-hardware-robust/v2-mild-r2/stage-c/checkpoint.pt",
+  "command_speed_mps": 2.2,
+  "duration_s": 10.0,
+  "final_absolute_heading_error_deg": {
+    "mean": 78.06403519401661,
+    "median": 69.26634856187916,
+    "p90": 153.97044094998287
+  },
+  "flight_events_per_second": 6.991552829742432,
+  "flight_fraction": 0.6431052684783936,
+  "forward_speed_mps": {
+    "clean_survivor_mean": 1.6127194166183472,
+    "mean": 1.6121970415115356,
+    "median": 1.6359825134277344,
+    "p10": 1.4503546953201294,
+    "p90": 1.7518620491027832
+  },
+  "initial_heading_displacement_speed_mps": {
+    "mean": 0.9829199910163879,
+    "median": 1.1384128332138062,
+    "p10": 0.1604430228471756
+  },
+  "max_tilt_deg_median": 16.76836371428601,
+  "mean_absolute_heading_error_deg": 43.80756975146533,
+  "mean_absolute_lateral_speed_mps": 0.1000186949968338,
+  "mean_absolute_yaw_rate_radps": 1.5083571672439575,
+  "num_envs": 256,
+  "survival_fraction": 0.984375,
+  "task_id": "Mjlab-Running-Flat-Backlash-MicroDuck"
+}

eval/stress-plain.json

@@ -0,0 +1,37 @@
+{
+  "absolute_lateral_displacement_m": {
+    "mean": 6.924654006958008,
+    "median": 7.288991451263428,
+    "p90": 11.336601257324219
+  },
+  "actor_observation_dim": 61,
+  "checkpoint": "/net/gcp/cloud-storage/europe-west/training/hannes/microduck_running/20260830-hardware-robust/v2-mild-r2/stage-c/checkpoint.pt",
+  "command_speed_mps": 2.2,
+  "duration_s": 10.0,
+  "final_absolute_heading_error_deg": {
+    "mean": 75.89308688708473,
+    "median": 66.99265371926279,
+    "p90": 149.8832694110256
+  },
+  "flight_events_per_second": 7.019308090209961,
+  "flight_fraction": 0.6557417511940002,
+  "forward_speed_mps": {
+    "clean_survivor_mean": 1.636256456375122,
+    "mean": 1.6349653005599976,
+    "median": 1.6558061838150024,
+    "p10": 1.4708143472671509,
+    "p90": 1.7637542486190796
+  },
+  "initial_heading_displacement_speed_mps": {
+    "mean": 1.0107059478759766,
+    "median": 1.1746344566345215,
+    "p10": 0.1442173719406128
+  },
+  "max_tilt_deg_median": 16.412852367986844,
+  "mean_absolute_heading_error_deg": 43.2924983559655,
+  "mean_absolute_lateral_speed_mps": 0.10448182374238968,
+  "mean_absolute_yaw_rate_radps": 1.3874105215072632,
+  "num_envs": 512,
+  "survival_fraction": 0.98828125,
+  "task_id": "Mjlab-Running-Flat-MicroDuck"
+}

legacy/iteration-8749/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f
+size 4843445

legacy/iteration-8749/manifest.json

@@ -0,0 +1,103 @@
+{
+  "schema_version": 4,
+  "model_api": 1,
+  "name": "microduck-running",
+  "kind": "perpetual",
+  "obs_len": 61,
+  "action_len": 14,
+  "action_scale": 1.0,
+  "entry_pose": "standing",
+  "duration_s": null,
+  "status": "sim-only",
+  "description": "Forward-running policy for the stock MicroDuck on flat ground. This is checkpoint 8749 used in the run-into-mat video, not the later faster continuation.",
+  "command": {
+    "layout": "twist(3), head_pose(4), body_pose(6)",
+    "twist": [
+      "forward velocity command in m/s; trained through 1.65, recommended simulation evaluation point 1.80; not a hard speed limit",
+      "lateral velocity; use 0 (training range was only -0.02 to 0.02 m/s)",
+      "yaw rate; use 0 (training range was only -0.05 to 0.05 rad/s)"
+    ],
+    "head": "unused; zeros",
+    "body": "unused; zeros",
+    "idle": [
+      0,
+      0,
+      0
+    ],
+    "exact_zero_training_fraction": 0.03
+  },
+  "robot": {
+    "model": "microduck",
+    "hw_rev": 1,
+    "servos": "14x Dynamixel XL330",
+    "control_hz": 50,
+    "hardware_modifications_required": false
+  },
+  "training": {
+    "task_id": "Mjlab-Running-Flat-MicroDuck",
+    "algorithm": "PPO (rsl_rl)",
+    "repo": "Vottivott/microduck-playground",
+    "commit": "c5fcc50219fef01ac9931d0079c583ccbb29b689",
+    "upstream_base": "pollen-robotics/microduck_rl@d424a0c899f6b33cbd3daeb279913134349c0b63",
+    "run": "sweep-speed-r1",
+    "checkpoint_iteration": 8749,
+    "resumed_from": "highspeed-r6 iteration 6750",
+    "continuation_iterations": 2000,
+    "environment": {
+      "MICRODUCK_RUNNING_TARGET_MAX_SPEED": 1.65,
+      "MICRODUCK_RUNNING_SPEED_CAP": 1.8,
+      "MICRODUCK_RUNNING_ACTION_RATE_WEIGHT": -0.1,
+      "MICRODUCK_RUNNING_ENABLE_SYMMETRY": false,
+      "MICRODUCK_RUNNING_ENABLE_HEADING_FEEDBACK": false
+    },
+    "export": "scripts/export.py; observation normalizer baked into policy.onnx",
+    "source_commit": "c5fcc50219fef01ac9931d0079c583ccbb29b689",
+    "continuation_checkpoint": {
+      "path": "checkpoint.pt",
+      "sha256": "ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f",
+      "actor_exact_to_released_onnx": true,
+      "original_full_ppo_state": true
+    }
+  },
+  "eval": {
+    "sim_proxy": "mjlab / MuJoCo Warp with BAM XL330 M6 actuator model",
+    "command_speed_m_s": 1.8,
+    "num_envs": 256,
+    "duration_s": 8.0,
+    "survival_fraction": 0.9921875,
+    "mean_body_forward_speed_m_s": 1.6072514057159424,
+    "mean_initial_heading_displacement_speed_m_s": 1.3276939392089844,
+    "mean_absolute_heading_error_deg": 25.215665057096395,
+    "mean_absolute_lateral_displacement_m": 3.8727223873138428,
+    "flight_fraction": 0.6572428941726685,
+    "known_limits": "Heading drift is substantial. At a 2.00 m/s extrapolated command, mean body-forward speed was 1.6452 m/s and survival fell to 0.9336. Never tested on hardware."
+  },
+  "sim2real": {
+    "retained": [
+      "BAM XL330 M6 voltage control, back-EMF, current limiting, battery 6.5-8.2 V, and load-dependent sag",
+      "3-6 physics-step actuator delay, 0-1-step IMU delay, and one-control-step joint-velocity lag",
+      "foot friction 0.7-1.3",
+      "trunk mass/inertia +/-5%, trunk/head CoM +/-3 mm, armature +/-10%, BAM joint friction +/-10%",
+      "encoder bias +/-0.015 rad, IMU misalignment up to 6 degrees, and actor observation noise"
+    ],
+    "relaxed": [
+      "flat terrain only",
+      "random push disturbances removed",
+      "non-backlash robot model",
+      "CoM/head-CoM and head/body pose-command widening curricula removed",
+      "weaker posture, angular-motion, foot-slip, and action-rate regularization; head/body pose tracking disabled",
+      "motor-gain randomization and randomized initial base tilt disabled in inherited base recipe"
+    ],
+    "hardware_validated": false
+  },
+  "media": {
+    "preview": "media/preview.mp4",
+    "run_into_mat": "media/run-into-mat.mp4",
+    "run_into_mat_note": "Same policy checkpoint; obstacle, collision augmentation, camera treatment, motor cutoff, and sparks are scripted demo elements. The policy has no obstacle perception.",
+    "audio": false
+  },
+  "source_artifact_sha256": "b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a",
+  "artifacts": {
+    "policy_onnx_sha256": "b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a"
+  }
+}

legacy/iteration-8749/media/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69ce8b5f8855ac76d2413d90b176e770d22d22a72766dcc2dacf34ccbded58d6
+size 3133555

legacy/iteration-8749/policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a
+size 793819

lineage/iteration-11748/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e79df200d8bb19b139ee1e74e6910c62434d57e792f632b66b10a73266894eb
+size 4843527

lineage/iteration-11748/manifest.json

@@ -0,0 +1,13 @@
+{
+  "checkpoint_iteration": 11748,
+  "role": "speed-focused parent of the robustified default",
+  "command_speed_m_s": 2.2,
+  "mean_body_forward_speed_m_s": 1.6873,
+  "survival_fraction": 0.9883,
+  "num_envs": 1024,
+  "duration_s": 10.0,
+  "policy_onnx_sha256": "ccd6757b53b591081fdf75dd3effe9a5743f5fac0c2a513fa48dfd5a32231b65",
+  "checkpoint_pt_sha256": "1e79df200d8bb19b139ee1e74e6910c62434d57e792f632b66b10a73266894eb",
+  "preview_mp4_sha256": "256ad675984d544204076986a7e09ea3a66c6070ec15ffb52921212e685e898c",
+  "evaluations_tar_sha256": "a81a468c789b890f935e14f478f992c632b01eeb25179da7d3ba252d663625be"
+}

lineage/iteration-11748/policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccd6757b53b591081fdf75dd3effe9a5743f5fac0c2a513fa48dfd5a32231b65
+size 793831

lineage/iteration-11748/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:256ad675984d544204076986a7e09ea3a66c6070ec15ffb52921212e685e898c
+size 3244149

manifest.json

@@ -1,5 +1,5 @@
 {
-  "schema_version": 4,
+  "schema_version": 5,
   "model_api": 1,
   "name": "microduck-running",
   "kind": "perpetual",
@@ -8,22 +8,18 @@
   "action_scale": 1.0,
   "entry_pose": "standing",
   "duration_s": null,
-  "status": "sim-only",
-  "description": "Forward-running policy for the stock MicroDuck on flat ground. This is checkpoint 8749 used in the run-into-mat video, not the later faster continuation.",
+  "status": "simulation-only-hardware-candidate",
+  "description": "Robustified iteration-12195 forward-running policy for the stock MicroDuck on flat ground. It is not hardware validated.",
   "command": {
     "layout": "twist(3), head_pose(4), body_pose(6)",
     "twist": [
-      "forward velocity command in m/s; trained through 1.65, recommended simulation evaluation point 1.80; not a hard speed limit",
+      "forward velocity command in m/s; curriculum trained through 2.2; evaluated at 2.2; not a hard speed limit",
       "lateral velocity; use 0 (training range was only -0.02 to 0.02 m/s)",
       "yaw rate; use 0 (training range was only -0.05 to 0.05 rad/s)"
     ],
     "head": "unused; zeros",
     "body": "unused; zeros",
-    "idle": [
-      0,
-      0,
-      0
-    ],
+    "idle": [0, 0, 0],
     "exact_zero_training_fraction": 0.03
   },
   "robot": {
@@ -37,67 +33,91 @@
     "task_id": "Mjlab-Running-Flat-MicroDuck",
     "algorithm": "PPO (rsl_rl)",
     "repo": "Vottivott/microduck-playground",
-    "commit": "c5fcc50219fef01ac9931d0079c583ccbb29b689",
+    "commit": "828d950134e29a8d04cbb51720a22c8729047fb7",
     "upstream_base": "pollen-robotics/microduck_rl@d424a0c899f6b33cbd3daeb279913134349c0b63",
-    "run": "sweep-speed-r1",
-    "checkpoint_iteration": 8749,
-    "resumed_from": "highspeed-r6 iteration 6750",
-    "continuation_iterations": 2000,
-    "environment": {
-      "MICRODUCK_RUNNING_TARGET_MAX_SPEED": 1.65,
-      "MICRODUCK_RUNNING_SPEED_CAP": 1.8,
+    "checkpoint_iteration": 12195,
+    "resumed_from": "iteration 11748",
+    "num_envs": 4096,
+    "seed": 42,
+    "common_environment": {
+      "MICRODUCK_RUNNING_TARGET_MAX_SPEED": 2.2,
+      "MICRODUCK_RUNNING_SPEED_CAP": 2.4,
       "MICRODUCK_RUNNING_ACTION_RATE_WEIGHT": -0.1,
+      "MICRODUCK_RUNNING_HIGH_SPEED_STAGE_INTERVAL": 750,
+      "MICRODUCK_RUNNING_FORWARD_PROGRESS_WEIGHT": 5.0,
       "MICRODUCK_RUNNING_ENABLE_SYMMETRY": false,
-      "MICRODUCK_RUNNING_ENABLE_HEADING_FEEDBACK": false
+      "MICRODUCK_RUNNING_ENABLE_HEADING_FEEDBACK": false,
+      "foot_friction_range": [0.7, 1.3],
+      "push_interval_s": [3.0, 6.0]
     },
+    "robustification_stages": [
+      {"updates": 100, "push_m_s": 0.03, "trunk_com_m": 0.003, "head_com_m": 0.003, "initial_tilt_deg": 1.0},
+      {"updates": 150, "push_m_s": 0.06, "trunk_com_m": 0.005, "head_com_m": 0.005, "initial_tilt_deg": 1.5},
+      {"updates": 200, "push_m_s": 0.1, "trunk_com_m": 0.008, "head_com_m": 0.006, "initial_tilt_deg": 2.0}
+    ],
     "export": "scripts/export.py; observation normalizer baked into policy.onnx",
-    "source_commit": "c5fcc50219fef01ac9931d0079c583ccbb29b689",
     "continuation_checkpoint": {
       "path": "checkpoint.pt",
-      "sha256": "ce7707c985d84affda1befec172a8c0a72fafaa4d3476a26bec11dde9199907f",
+      "sha256": "052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973",
       "actor_exact_to_released_onnx": true,
       "original_full_ppo_state": true
     }
   },
   "eval": {
     "sim_proxy": "mjlab / MuJoCo Warp with BAM XL330 M6 actuator model",
-    "command_speed_m_s": 1.8,
-    "num_envs": 256,
-    "duration_s": 8.0,
-    "survival_fraction": 0.9921875,
-    "mean_body_forward_speed_m_s": 1.6072514057159424,
-    "mean_initial_heading_displacement_speed_m_s": 1.3276939392089844,
-    "mean_absolute_heading_error_deg": 25.215665057096395,
-    "mean_absolute_lateral_displacement_m": 3.8727223873138428,
-    "flight_fraction": 0.6572428941726685,
-    "known_limits": "Heading drift is substantial. At a 2.00 m/s extrapolated command, mean body-forward speed was 1.6452 m/s and survival fell to 0.9336. Never tested on hardware."
+    "command_speed_m_s": 2.2,
+    "duration_s": 10.0,
+    "warmup_s": 1.0,
+    "results": {
+      "nominal_plain": {"num_envs": 512, "speed_m_s": 1.6510932445526123, "survival": 0.9921875},
+      "stress_plain": {"num_envs": 512, "speed_m_s": 1.6349653005599976, "survival": 0.98828125},
+      "nominal_backlash": {"num_envs": 256, "speed_m_s": 1.6362495422363281, "survival": 0.98828125},
+      "stress_backlash": {"num_envs": 256, "speed_m_s": 1.6121970415115356, "survival": 0.984375},
+      "high_grip_stress": {"num_envs": 512, "speed_m_s": 1.6424154043197632, "survival": 0.966796875}
+    },
+    "stress": {"push_m_s": 0.1, "trunk_com_m": 0.01, "head_com_m": 0.006, "initial_tilt_deg": 2.0},
+    "known_limits": "Heading and lateral drift remain substantial. Never tested on hardware."
+  },
+  "onnx_contract": {
+    "observation_normalizer_embedded": true,
+    "training_action_clipping": false,
+    "onnx_output_clip": false,
+    "checkpoint_parity_max_abs_error": 0.00000762939453125
   },
   "sim2real": {
     "retained": [
       "BAM XL330 M6 voltage control, back-EMF, current limiting, battery 6.5-8.2 V, and load-dependent sag",
       "3-6 physics-step actuator delay, 0-1-step IMU delay, and one-control-step joint-velocity lag",
       "foot friction 0.7-1.3",
-      "trunk mass/inertia +/-5%, trunk/head CoM +/-3 mm, armature +/-10%, BAM joint friction +/-10%",
+      "trunk mass/inertia +/-5%, trunk CoM +/-8 mm, head CoM +/-6 mm, armature +/-10%, and BAM joint friction +/-10% in the final training stage",
+      "planar velocity pushes up to +/-0.10 m/s and initial pitch/roll up to +/-2 degrees in the final training stage",
       "encoder bias +/-0.015 rad, IMU misalignment up to 6 degrees, and actor observation noise"
     ],
     "relaxed": [
       "flat terrain only",
-      "random push disturbances removed",
-      "non-backlash robot model",
-      "CoM/head-CoM and head/body pose-command widening curricula removed",
+      "trained on the ordinary non-backlash robot model; backlash is cross-evaluation only",
+      "head/body pose-command widening curricula removed",
       "weaker posture, angular-motion, foot-slip, and action-rate regularization; head/body pose tracking disabled",
-      "motor-gain randomization and randomized initial base tilt disabled in inherited base recipe"
+      "motor-gain randomization remains disabled in the inherited recipe"
     ],
     "hardware_validated": false
   },
+  "versions": {
+    "default": {"iteration": 12195, "policy": "policy.onnx", "checkpoint": "checkpoint.pt", "preview": "media/preview.mp4"},
+    "speed_parent": {"iteration": 11748, "directory": "lineage/iteration-11748"},
+    "legacy_mat_policy": {"iteration": 8749, "directory": "legacy/iteration-8749"}
+  },
   "media": {
     "preview": "media/preview.mp4",
     "run_into_mat": "media/run-into-mat.mp4",
-    "run_into_mat_note": "Same policy checkpoint; obstacle, collision augmentation, camera treatment, motor cutoff, and sparks are scripted demo elements. The policy has no obstacle perception.",
+    "run_into_mat_policy_iteration": 8749,
+    "run_into_mat_note": "The mat, collision augmentation, camera treatment, motor cutoff, and sparks are scripted demo elements. The policy has no obstacle perception.",
     "audio": false
   },
-  "source_artifact_sha256": "b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a",
+  "source_artifact_sha256": "007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb",
   "artifacts": {
-    "policy_onnx_sha256": "b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a"
+    "policy_onnx_sha256": "007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb",
+    "checkpoint_pt_sha256": "052f6df6683fdae83deb369b7f2c7d13e45f87ed2907adee6e23b1ebf9dfe973",
+    "preview_mp4_sha256": "942f0802cc0c6795769bd99a9991f866fee682890ea9ea6d71077f7d4eb5492b"
   }
 }

media/preview.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:69ce8b5f8855ac76d2413d90b176e770d22d22a72766dcc2dacf34ccbded58d6
-size 3133555
+oid sha256:942f0802cc0c6795769bd99a9991f866fee682890ea9ea6d71077f7d4eb5492b
+size 3487586

policy.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b4e285fb3671f52c28f671bd4733a4e11fa17c21f40d25f52eaf2614f9e2e02a
-size 793819
+oid sha256:007707dd7779b2756ded67c58b2e9f94fe5071794a48c2b5a20d5f8d841efbeb
+size 793744