Zygot is a data lab building egocentric datasets for embodied AI. We license first-person manipulation data to frontier labs training vision-language-action policies, world models and video models.

Teleoperation is expensive to scale, and internet video lacks the signals a policy needs to learn action. Egocentric human data sits between the two: the viewpoint of an embodied agent, the diversity of real human activity, and the 3D structure required to ground it.

Data

Multimodal and synchronized. Head-mounted RGB and stereo depth with calibrated intrinsics and extrinsics, 3D bimanual hand pose (MANO), eye gaze, 6-DoF head trajectory and IMU, hardware-synced to a single clock.

Long-horizon and in the wild. Multi-step tasks across homes, factories and warehouses, annotated with step-level action segments, language narration, object identities and hand-object contact.

Delivery

Built for training pipelines. Delivered as LeRobot, RLDS, WebDataset or MCAP, with hand trajectories ready to retarget onto humanoid and dexterous embodiments.

Off the shelf or to spec. License existing corpora, or commission capture against your own task taxonomy, environments and hardware.

Rights

Clean by construction. Every clip carries contributor consent, location releases, PII redaction and full provenance. Licenses cover pretraining, post-training and evaluation of commercial models, on non-exclusive or exclusive terms.

Team

We are researchers in robot learning, egocentric vision and 3D perception. We build the datasets used by frontier labs.

Contact

Samples and specifications are available on request at .