phydata

Custom training data for physical AI.

Tell us the capability your model is missing. We design and run the collection, capture real demonstrations, edge cases, and task signals, then deliver synchronized training data in your schema.

See how it works

Not more video. phydata keeps what decided the task.

Motion, conditions, tool state, decisions, failures, and outcomes, recorded together on one clock.

Built by people from

Research and engineering backgrounds spanning MIT and Waterloo, with founding engineering experience at a YC S26 company.

We capture the data your physical AI is missing.

For robotics labs and model teams, phydata scopes the task, conditions, failures, outcomes, and required signals, then builds the collection path around the model gap. Our sensing roadmap adds force, pressure, and other tactile signals where video is not enough.

See collection framework

  1. Define the capability, conditions, signals, and success criteria.

    The brief decides which streams exist before anyone records.

  2. Run real demonstrations and retain failures, corrections, and recoveries.

    Each stream records on its own clock. The re‑grip stays in.

  3. Synchronize every stream and link each derived label to its source evidence.

    One timeline. Select an event to see the frame it came from.

  4. Provide validated training data in the model team's schema.

    Every stream becomes a file your loader already understands.

Plane a dowel round

Brief

Capability
Plane a dowel round by hand
Conditions
Hardwood, grain reversals, either hand
Signals
Overhead video, hand motion, tool angle; contact force when the sensor is fitted
Success
An even profile with no tear-out
Overhead videoPlanned+0 ms60 fpsoverhead_rgb.mp4
Hand motionPlanned+86 ms120 Hzhand_motion.parquet
Tool anglePlanned−58 ms200 Hztool_angle.parquet
Contact forceRoadmapPlannedPlannedcontact_force.parquet
EventsDerivedevents.json

Define the capability, conditions, signals, and success criteria.

The brief decides which streams exist before anyone records.

Roadmap

Teach a machine the task only you know.

Today, phydata runs the collection. Next, we plan to develop a lower-cost human-side capture toolkit based on our collection workflow, so teams can record private, unusual, or highly specific tasks where they happen.

In development. Capture tools for people, not robot parts.

A stubborn valve. A private assembly step. A household routine too specific to expect in a public dataset.

Concept drawing

010203

Domains we design collection for

Data collected.

One episode: every stream on one clock, in your schema.

Field CaptureReal work, instrumented where it happens

What we learn from real collection, written down.

All research

Every physical AI model has a blind spot.

Continuous collection builds coverage. Targeted collection closes the gaps.

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.