Physical AI datasets. Built to your spec.
We design and run human-demonstration capture for robotics and physical AI: one egocentric view, up to four exocentric cameras and optional wrist IMUs, captured by trained operators and delivered training-ready.
One spec. Three ways to collect.
POD capture
Structured tasks in a repeatable room, with one egocentric camera and up to four exocentric views placed around what the model needs to see.
Network capture
Trained people across real homes, workshops and everyday places, for when variety matters more than one controlled room.
Robot-arm demonstrations
We are building a network of operators with arms, so future projects can collect teleoperated demonstrations for ACT and related policy training.
Only what earns its place.
More hardware is not automatically better data. Each view and sensor is chosen for the signal the model needs.
- Views
- One head-mounted camera and up to four fixed exocentric cameras. The count follows the spec, not a bundle.
- Motion
- Optional wrist IMUs when the project needs motion beyond visual hand and body tracking.
- Sync and calibration
- Flash and audio sync with calibrated cameras today, declared per session.
- Hardware
- iPhone with Record3D and standard video now. Depth, more ego devices and robot embodiments are added when the signal justifies them.
Pixels are only the first layer.
Processing can add geometry, hands, bodies, objects, contact, action and quality signals. Every delivery separates what was measured from what was estimated, so you know what to trust.
A dataset you can trace and rerun.
Write the capture spec
Tasks, environments, viewpoints, hardware, participants, signals, acceptance thresholds and delivery format, agreed with your team.
Choose the capture mode
The POD, distributed operators, or both, for the coverage the policy needs.
Train and capture
Operators get the task protocol, equipment guidance, consent terms and quality criteria before a session starts.
Process and score
Source media is kept, signals are added where the capture supports them, and every asset is scored against the brief.
Deliver with context
The agreed format with manifests, a dataset card, licensing, provenance and known limitations.
Data, plus the context to use it.
-
Video and synchronized views
Source media and normalized camera streams, organized by session and task.
-
Signals with provenance
Hand and body pose, detections, tracks, contact, camera pose, depth, IMU and action channels, where the capture and brief support them.
-
Training-ready exports
WebDataset, RLDS and LeRobot or GR00T-compatible layouts, scoped to your training stack.
-
Dataset card and QA
Acceptance results, measured versus estimated, licensing, limitations and asset provenance.
From human demonstrations to real-arm policies.
Our pipeline already exports LeRobot-compatible episodes and can train and evaluate ACT policies. The next step is operators who own arms and teleoperate tasks outside a single lab. This track is in buildout: we are recruiting arm owners and scoping pilots with teams that want to help shape the protocol.
Discuss an ACT pilotGet closer to the data.
Our public samples on Hugging Face (opens in a new tab)
Appliance maintenance
Tape removal, brushing and wiping from a chest-mounted view: six excerpts with source-frame mapping and approximate action intervals.
(opens Hugging Face in a new tab)
Kitchen manipulation
Onion prep, pepper slicing and dicing: three recordings with native RGB-D, device sensors and estimated annotations.
(opens Hugging Face in a new tab)Planning your dataset.
What should a dataset brief include?
The tasks, environments, participant profile, camera views, required signals, acceptance criteria and training format, and which signals must be measured rather than estimated.
POD or distributed capture?
The POD suits tasks that need a repeatable room and controlled camera placement. Distributed capture adds variety across homes, workshops and everyday places. The task decides, and projects can combine both.
Which formats and signals can you deliver?
Source video, synchronized views, QA results, dataset cards, licensing and provenance, in WebDataset, RLDS or LeRobot/GR00T-compatible layouts. Depth, pose, IMU and action channels depend on the hardware and the brief.
Can I take part as a capture operator?
The Operators Lab is the way in. Project fit depends on demand, location, skills and equipment, and joining does not guarantee paid work.