Every skilled action contains physical signals. Most of them are never captured.

Volume is not coverage.

A million easy clips can still miss the decision that makes a task succeed.

We map variation across people, objects, tools, environments, failures, and recoveries, then collect where the map is thin.

The result is not just footage. It is synchronized physical data that keeps signals, decisions, and outcomes together.

Collection follows the model.

When evaluation exposes a missing capability, we translate it into a field program with the exact task, conditions, sensing, edge cases, and success criteria.

Each program is shaped around the evidence that could change model behavior, not a generic clip target.

The loop repeats: train, evaluate, collect the missing reality, and retrain.

The phydata system

ScopeDefine the capability, conditions, signals, and success criteria.Purpose-built

CaptureRun real demonstrations and retain failures, corrections, and recoveries.In the wild

StructureSynchronize every stream and link each derived label to its source evidence.Traceable

DeliverProvide validated training data in the model team's schema.Model-ready

Current focus

Human task captureEmbodied AIActive

Contact-rich task dataManipulationResearch

Failure and recovery dataEvaluationOngoing

Human-worn capture toolsHardwareRoadmap

Seasonal field variationAgricultureScoping

Multimodal synchronizationInfrastructureActive

Tell us the capability your model is missing.

Describe the task, the setting, and what success looks like. We design the collection around it and deliver the data in your schema.