Training data for physical AI.

We operate our own robot fleet and collect multimodal training data from real-world deployments.

Customers

Sensor streams and delivery

Video, joint and motion sensor data recorded in sync, labeled by people and delivered in standard formats.

01
Imitation learning
Real robot runs labeled by outcome, with every state and action lined up with the video and joint data. Ready for behavior cloning and inverse RL.
02
Foundation model pre-training
Large amounts of data across many tasks and kinds of robot, with vision, language and action together, for pre-training general policies.
03
Human-robot interaction
People moving, filmed from their own point of view and from the outside at the same time, with 3D body pose for every frame.
04
Production deployment
Failures and edge cases from robots working in the real world, collected continuously for online learning and policy updates.

We collect on two kinds of robot: a humanoid for handling objects and a two-legged robot for walking.

A humanoid with 23 degrees of freedom and a gripper, used for pick-and-place, tool use and fine assembly.

A two-legged robot used to record walking, balancing and whole-body movement.

Our robot and human-motion data comes as three products.

Robot Teleoperation Demonstrations	Robots operated by people, recorded as synced video, robot actions and whether each task succeeded.
Robot Sensor Recordings	Joint, motion-sensor and audio recordings from several kinds of robot, synced to under a millisecond.
First-Person Human Motion	First-person video of people doing tasks, paired with full-body motion capture.

You can look at sample episodes before we scope a delivery. Full collections run on our own infrastructure and scale to petabytes.