Data types

Three paradigms of real-world robot training data

Real-robot teleoperation, human demonstration with handheld devices, and egocentric video. Each serves a different stage of training humanoid, mobile and arm-based robots, from VLA pre-training to embodiment-specific fine-tuning.

Data types

Three paradigms, one data pyramid.

Combine all three for breadth, scale and grounding on real robots.

7,000 h

Real-Robot Teleoperation

Operators teleoperate real robots through each task. Every episode pairs synchronized video with robot state, action trajectories and task labels. One package also includes contact-force and torque signals.

Best for: embodiment-specific fine-tuning, policy evaluation, contact-rich tasks.

80,000 h

Human Demonstrationwith handheld devices

Calibrated, synchronized three-camera capture (head + left wrist + right wrist, 1080p by default), end-effector pose and gripper-state trajectories, hand motion, task labels. Robot-agnostic and retargetable to many embodiments.

Best for: large-scale policy pre-training, cross-embodiment transfer, dexterous and deformable-object manipulation.

22,000 h

Egocentric Video

First-person video of real task processes, recorded step by step with task metadata. Mono or stereo capture.

Best for: world-model and VLA pre-training, task understanding, long-horizon planning priors.

Modalities by collection paradigm
Modality	Teleoperation	Human Demonstration	Egocentric
Video / Image		calibrated tri-camera	mono or stereo
Robot State
Motion / Trajectory
Force / Tactile	one selected package
Hand / Gripper Motion
Labels & Task Metadata

Teleoperation data for robot learning

Operators drive real robots through each task, so every episode carries the robot's own state and actions. Collected on 7 commercial platforms, including bipedal humanoids, wheeled humanoids and mobile dual-arm systems.

Human demonstration data

UMI-style handheld demonstration data: calibrated tri-camera capture with end-effector and gripper trajectories, robot-agnostic and retargetable. It is the largest paradigm in the catalog at 80,000 hours.

Egocentric video dataset for robotics

First-person video of real task processes with step-level task metadata, used for world-model and VLA pre-training and long-horizon planning priors.

Not sure which you need? Read teleoperation vs human demonstration vs egocentric video, or browse the catalog.

Evaluate the data in your own pipeline.

Tell us what you need. We reply within 3 business days.

Request a sample Send a data request