Data Curation for Physical AI

We turn video and sensor data into signals: actions, contact, trajectories and quality. Choose better data, train better models.

TRY IT ON YOUR DATA

Signals you can trust

understands what happens over time in video and telemetry. It is faster, cheaper and more accurate than frontier models, measured against human-verified clips.

REASSEMBLE: scenario mining

F1, mean of 10 scenarios

0.80
GPT-6 Astra0.32
Gemini 3.8 Flash0.14

Figure 1: Accuracy (F1) against human-verified clips on REASSEMBLE (Sliwowski et al., 2025): 149 robot assembly episodes, 13.3 hours of video and telemetry. GPT-6 Astra and Gemini 3.8 Flash watched every episode at 1 fps.

Know what's in your data. Then train on it.

Understand

Uncover signals: actions over time, contact and hand–object interactions, trajectories, and whether an example is reliable enough to learn from.

Search

Inspect any behavior, failure or scene with a simple description, linked to the exact moment in the source.

Curate

Select and filter training data in your own pipeline, or use signals as supervision targets. Versioned with each data release.