Physical AI

Data for physical AI, captured where the work happens.

Egocentric video, teleoperation demonstrations, and manipulation data from real factories, clinics, kitchens, and homes. Sourced from existing holders or captured to your spec.

Modalities

Egocentric / POV video

First-person footage of skilled work as it actually happens — machine tending, assembly, food prep, patient handling. Long, uncut episodes give world models the context that staged clips never carry, and each session is segmented into action-level spans for downstream VLA training.

1080p–4K30/60 fpsAction segmentationHead-mounted

Teleoperation demonstrations

Operator-driven episodes recorded on real arms and grippers, with synchronized observation and action streams. Collected across multiple robot bodies so cross-embodiment policies and imitation learning pipelines have more than one morphology to generalize from.

Leader–followerCross-embodimentEpisode metadataImitation learning

Manipulation trajectories

Dense trajectories for grasping, insertion, tool use, and bimanual tasks, with joint states, end-effector poses, and success labels per episode. Structured for sim-to-real transfer, so policies trained in simulation can be fine-tuned on matched real captures.

Joint statesEE poseSuccess labelsSim-to-real

Sensor fusion (IMU, depth, force)

Time-aligned IMU, depth, and force/torque streams recorded alongside video. Hardware-synced timestamps keep every modality on one clock, so trajectories stay usable when contact-rich behaviour matters more than pixels.

IMU 200 HzStereo/ToF depthForce/torqueHardware sync

Human motion & hand pose

Full-body and hand keypoint tracks derived from multi-view capture of real workers. Used as human priors for world models and as demonstration sources when no robot is in the loop, with action segmentation labels attached to each episode.

21-pt hand poseBody keypointsMulti-viewWorld models

Environments

Factory floorMachine shopWarehouseClinicRetailKitchenResidentialAgriculture

Delivery formats

LeRobotRLDSHDF5MCAPROS bagMP4 + JSON

Provenance you can defend.

Consent per jurisdiction

Every person on camera signs a consent form in their own language, written against the privacy rules of the country the site sits in, and is paid for taking part.

Commercial licensing

Named-purpose commercial licences agreed in writing before capture begins, with the recipient identified up front. No open-market resale.

De-identification

Faces, licence plates, badges, machine and computer screens, and voices are blurred or removed to the spec agreed during the site walkthrough.

Audit-ready records

Consent forms, site agreements, capture logs, and delivery approvals are retained per episode so provenance can be reconstructed end to end.

How it works

01

Brief

Tell us the task, embodiment, environment, and scale. The more specific the action space, the faster we move.

02

Source or capture

We check existing holders first. If the episodes don't exist, we schedule capture at partner sites to your spec.

03

Verify and deliver

QC on sync, labels, and coverage, then delivery in LeRobot, RLDS, HDF5, or your own schema with full chain of custody.

Tell us what you need captured.

Send the task, the embodiment, and the scale. We'll come back within 48 hours with what exists today and what we can capture.

Talk to us