Upload folder using huggingface_hub

README.md

@@ -0,0 +1,26 @@
+---
+license: mit
+task_categories:
+- robotics
+language:
+- en
+tags:
+- libero-plus
+- opentau
+- pi0.5
+- robotics
+- benchmark
+- summary
+pretty_name: pi0.5 LIBERO-plus Perturbation Summary
+---
+
+# pi0.5 LIBERO-plus Perturbation Summary
+
+This dataset stores the summary report for six completed LIBERO-plus perturbation evaluations of `TensorAuto/tPi0.5-libero`.
+
+Files:
+
+- `pi05_libero_plus_six_perturb_hf_report.md`: human-readable report with Hugging Face links, success rates, and short analysis.
+- `pi05_libero_plus_six_perturb_hf_report.json`: machine-readable summary.
+
+The failure-grid videos and per-category metadata are stored in the linked per-perturbation datasets.

pi05_libero_plus_six_perturb_hf_report.json

@@ -0,0 +1,74 @@
+{
+  "micro_success": 0.8114975845410628,
+  "macro_success": 0.8254923800827516,
+  "total_successes": 8399,
+  "total_episodes": 10350,
+  "rows": [
+    {
+      "title": "Camera Viewpoints",
+      "repo": "pi05-libero-plus-camera-view-failures",
+      "tasks": 419,
+      "episodes": 2095,
+      "successes": 1197,
+      "failures": 898,
+      "success_rate": 0.5713603818615752,
+      "grid_videos": 55,
+      "note": "Most brittle among the completed local categories; camera pose changes strongly affect visual grounding."
+    },
+    {
+      "title": "Robot Initial States",
+      "repo": "pi05-libero-plus-robot-initial-failures",
+      "tasks": 393,
+      "episodes": 1965,
+      "successes": 1430,
+      "failures": 535,
+      "success_rate": 0.727735368956743,
+      "grid_videos": 32,
+      "note": "Moderate robustness; failures often come from approach pose and grasp recovery sensitivity."
+    },
+    {
+      "title": "Light Conditions",
+      "repo": "pi05-libero-plus-light-conditions-failures",
+      "tasks": 274,
+      "episodes": 1370,
+      "successes": 1209,
+      "failures": 161,
+      "success_rate": 0.8824817518248175,
+      "grid_videos": 10,
+      "note": "Highest success among completed local categories; illumination shifts are less damaging than geometry/language changes."
+    },
+    {
+      "title": "Background Textures",
+      "repo": "pi05-libero-plus-background-textures-failures",
+      "tasks": 289,
+      "episodes": 1445,
+      "successes": 1277,
+      "failures": 168,
+      "success_rate": 0.8837370242214533,
+      "grid_videos": 11,
+      "note": "Texture/domain shift stresses visual appearance; failures are useful for checking whether policy overfits tabletop/background cues."
+    },
+    {
+      "title": "Language Instructions",
+      "repo": "pi05-libero-plus-language-instructions-failures",
+      "tasks": 383,
+      "episodes": 1915,
+      "successes": 1841,
+      "failures": 74,
+      "success_rate": 0.9613577023498695,
+      "grid_videos": 5,
+      "note": "Language paraphrase/instruction shift directly tests semantic grounding; low score usually indicates instruction following and target disambiguation errors."
+    },
+    {
+      "title": "Objects Layout",
+      "repo": "pi05-libero-plus-objects-layout-failures",
+      "tasks": 312,
+      "episodes": 1560,
+      "successes": 1445,
+      "failures": 115,
+      "success_rate": 0.9262820512820513,
+      "grid_videos": 8,
+      "note": "Object layout changes stress spatial search, grasp approach, and collision recovery; failures often expose planning rather than pure perception issues."
+    }
+  ]
+}

pi05_libero_plus_six_perturb_hf_report.md

@@ -0,0 +1,29 @@
+# pi0.5 LIBERO-plus Perturbation Result Summary
+
+This report combines the three earlier local perturbation uploads with the three `wzxuan` perturbation uploads prepared from `/data/wangjinqiu/tta-memory/libero_plus_eval`.
+
+## Overall
+
+- Categories summarized: `6`
+- Micro success: `8399/10350` = `81.15%`
+- Macro success across categories: `82.55%`
+- Policy: `TensorAuto/tPi0.5-libero`
+- Benchmark: LIBERO-plus `libero_10`, 5 episodes per task, seeds `1000..1004`
+
+## Hugging Face Datasets
+
+| Perturbation | HF dataset | Tasks | Episodes | Success | Failure grids | Short analysis |
+|---|---|---:|---:|---:|---:|---|
+| Camera Viewpoints | [pi05-libero-plus-camera-view-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-camera-view-failures) | 419 | 2095 | 1197/2095 (57.14%) | 55 | Most brittle among the completed local categories; camera pose changes strongly affect visual grounding. |
+| Robot Initial States | [pi05-libero-plus-robot-initial-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-robot-initial-failures) | 393 | 1965 | 1430/1965 (72.77%) | 32 | Moderate robustness; failures often come from approach pose and grasp recovery sensitivity. |
+| Light Conditions | [pi05-libero-plus-light-conditions-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-light-conditions-failures) | 274 | 1370 | 1209/1370 (88.25%) | 10 | Highest success among completed local categories; illumination shifts are less damaging than geometry/language changes. |
+| Background Textures | [pi05-libero-plus-background-textures-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-background-textures-failures) | 289 | 1445 | 1277/1445 (88.37%) | 11 | Texture/domain shift stresses visual appearance; failures are useful for checking whether policy overfits tabletop/background cues. |
+| Language Instructions | [pi05-libero-plus-language-instructions-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-language-instructions-failures) | 383 | 1915 | 1841/1915 (96.14%) | 5 | Language paraphrase/instruction shift directly tests semantic grounding; low score usually indicates instruction following and target disambiguation errors. |
+| Objects Layout | [pi05-libero-plus-objects-layout-failures](https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-objects-layout-failures) | 312 | 1560 | 1445/1560 (92.63%) | 8 | Object layout changes stress spatial search, grasp approach, and collision recovery; failures often expose planning rather than pure perception issues. |
+
+## Caveats
+
+- Success is the LIBERO/OpenTau task-completion predicate from `eval_info.json`; videos are qualitative diagnostics.
+- Failed-task reruns are excluded from success-rate computation.
+- The three earlier local datasets and the three `wzxuan` datasets were prepared on different machines, but use the same policy, task suite, 5-episode protocol, and seed schedule.
+- `Sensor Noise` is not included here because the downloaded `wzxuan` payload did not contain a failure-grid report for that category.