First Physics Audit of Open-X Embodiment — 216 Episodes, 78.1% Pass Rate

Research
1

# First Physics Audit of Open-X Embodiment — 216 Episodes, 78.1% Pass Rate

I built a tool that applies biomechanical physics laws to sensor data before training.

No ML. No learned classifier. Just equations — F=ma coupling, rigid-body kinematics,

jerk bounds, Hurst persistence.

I ran it on RoboTurk from Open-X Embodiment. Here is what came out.

-–

## The Audit

**Dataset:** RoboTurk (Open-X Embodiment, Stanford)

**Episodes:** 216 human-teleoperated demonstrations

**Windows certified:** 1,143

| Tier | Count | % |

|------|-------|—|

| GOLD | 284 | 24.8% |

| SILVER | 609 | 53.3% |

| BRONZE | 242 | 21.2% |

| REJECTED | 8 | 0.7% |

**Pass rate (GOLD+SILVER): 78.1%**

**Top failing law: `imu_internal_consistency` — 32.4% of windows**

-–

## What the Finding Means

`imu_internal_consistency` checks that translational acceleration and rotational

acceleration are physically coupled — as they are in real human motion.

In RoboTurk, `world_vector` (translation) and `rotation_delta` (rotation) are

commanded through separate channels in the smartphone teleoperation interface.

They have different latencies. S2S detects this mismatch.

This is not a bug in the data. It is a measurable property of the teleoperation

interface — and S2S quantifies it. 32.4% of windows have translational and

rotational commands that are physically inconsistent with each other.

For robot training: a model trained on these windows learns motion where the

hand translation and wrist rotation are decoupled. That is not how humans move.

-–

## Comparison to Real Human IMU

| Dataset | Pass Rate | Top Law |

|---------|-----------|---------|

| NinaPro DB5 (real human, 2000Hz) | 100% SILVER | none |

| RoboTurk (teleoperation, 15Hz) | 78.1% | imu_consistency 32.4% |

Real human IMU passes everything. Teleoperation data has a measurable quality gap.

-–

## Reproduce It

```bash

pip install s2s-certify

git clone GitHub - timbo4u1/S2S: Physics certification for robot training data. Checks 11 biomechanical laws before your model trains. · GitHub

cd S2S && python3 certify_roboturk.py

```

Full audit data: Scan2s/s2s-certified-motion · Datasets at Hugging Face

-–

This is the first physics audit of Open-X Embodiment I am aware of.

If anyone has run similar analysis on other Open-X subsets I would like to know.

The tool works on any IMU/EMG dataset. Zero dependencies. Pure Python.

1 Like