{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiakmxjmb342gj5sx3vbw2hzg65u5suwriq7df75ppfqo7jd3rg43e",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mgt5niw5jvb2"
  },
  "path": "/t/first-physics-audit-of-open-x-embodiment-216-episodes-78-1-pass-rate/174205#post_1",
  "publishedAt": "2026-03-12T00:10:02.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "GitHub - timbo4u1/S2S: Physics certification for robot training data. Checks 11 biomechanical laws before your model trains. · GitHub",
    "Scan2s/s2s-certified-motion · Datasets at Hugging Face"
  ],
  "textContent": "# First Physics Audit of Open-X Embodiment — 216 Episodes, 78.1% Pass Rate\n\nI built a tool that applies biomechanical physics laws to sensor data before training.\n\nNo ML. No learned classifier. Just equations — F=ma coupling, rigid-body kinematics,\n\njerk bounds, Hurst persistence.\n\nI ran it on RoboTurk from Open-X Embodiment. Here is what came out.\n\n-–\n\n## The Audit\n\n**Dataset:** RoboTurk (Open-X Embodiment, Stanford)\n\n**Episodes:** 216 human-teleoperated demonstrations\n\n**Windows certified:** 1,143\n\n| Tier | Count | % |\n\n|------|-------|—|\n\n| GOLD | 284 | 24.8% |\n\n| SILVER | 609 | 53.3% |\n\n| BRONZE | 242 | 21.2% |\n\n| REJECTED | 8 | 0.7% |\n\n**Pass rate (GOLD+SILVER): 78.1%**\n\n**Top failing law: `imu_internal_consistency` — 32.4% of windows**\n\n-–\n\n## What the Finding Means\n\n`imu_internal_consistency` checks that translational acceleration and rotational\n\nacceleration are physically coupled — as they are in real human motion.\n\nIn RoboTurk, `world_vector` (translation) and `rotation_delta` (rotation) are\n\ncommanded through separate channels in the smartphone teleoperation interface.\n\nThey have different latencies. S2S detects this mismatch.\n\nThis is not a bug in the data. It is a measurable property of the teleoperation\n\ninterface — and S2S quantifies it. 32.4% of windows have translational and\n\nrotational commands that are physically inconsistent with each other.\n\nFor robot training: a model trained on these windows learns motion where the\n\nhand translation and wrist rotation are decoupled. That is not how humans move.\n\n-–\n\n## Comparison to Real Human IMU\n\n| Dataset | Pass Rate | Top Law |\n\n|---------|-----------|---------|\n\n| NinaPro DB5 (real human, 2000Hz) | 100% SILVER | none |\n\n| RoboTurk (teleoperation, 15Hz) | 78.1% | imu_consistency 32.4% |\n\nReal human IMU passes everything. Teleoperation data has a measurable quality gap.\n\n-–\n\n## Reproduce It\n\n```bash\n\npip install s2s-certify\n\ngit clone GitHub - timbo4u1/S2S: Physics certification for robot training data. Checks 11 biomechanical laws before your model trains. · GitHub\n\ncd S2S && python3 certify_roboturk.py\n\n```\n\nFull audit data: Scan2s/s2s-certified-motion · Datasets at Hugging Face\n\n-–\n\nThis is the first physics audit of Open-X Embodiment I am aware of.\n\nIf anyone has run similar analysis on other Open-X subsets I would like to know.\n\nThe tool works on any IMU/EMG dataset. Zero dependencies. Pure Python.",
  "title": "First Physics Audit of Open-X Embodiment — 216 Episodes, 78.1% Pass Rate"
}