Data Types

Three core types of robot training data

Different capture methods serve different training paradigms — from egocentric perception to scalable hand-held demonstrations to real-robot data fully aligned with your deployment embodiment.

Egocentric Data

Egocentric

Head-, chest- or glasses-mounted first-person data with RGB, depth, IMU and hand trajectories — a real human perspective for vision-language-action models.

Best for: VLA · visual understanding · task planning

Browse 4 datasets

Universal Manipulation Interface

UMI Hand-held

Embodiment-free human demonstrations captured with hand-held devices: 6D poses, end-effector trajectories and grasp states, for low-cost skill collection at scale.

Best for: Behavior cloning · diffusion policy · cross-embodiment transfer

Browse 3 datasets

Teleoperation Data

Teleoperation

Multi-camera video plus full qpos/qvel/action signals from real-robot teleoperation — high-value data fully aligned with the deployment embodiment.

Best for: ACT · diffusion policy · real-robot manipulation

Browse 3 datasets

Not sure which data type fits your model?

Enter the market Talk to us