./data
The highest-quality humanoid training data.
For labs and robotics teams: real teleoperation episodes, captured at scale by certified human operators.
01THE BOTTLENECK
Humanoids learn from three data sources. Teleop is the only one with the signal to teach contact-rich manipulation — and the only one that doesn't scale on its own.
Egocentric video
scales · weak control signal
scale
signal
Synthetic
scales · sim-to-real gap
scale
signal
Teleoperation
highest signal · hard to scale
scale
signal
02WHAT WE CAPTURE
Every datapoint logged per episode — synchronized, timestamped, and labeled. Captured on a 29-DoF humanoid (Unitree G1) under whole-body teleoperation.
vision
- head cam
- stereo RGB · 640×480
- wrist cams
- RGB + stereo IR · ×2
- frame rate
- 30 fps · synced
- intrinsics
- fx/fy · ppx/ppy
- distortion
- brown-conrady · 5-coeff
- extrinsics
- R + T · per sensor
- rectification
- R/P/Q · 63mm baseline
state · action · labels
- joints
- 29-DoF · pos/vel/torque
- end-effector
- 6-DoF state + action
- gripper
- trigger + squeeze · L/R
- dex hands
- per-finger · L/R
- imu
- orientation + accel
- odometry
- base pose + mode
- subtasks
- 161 labels · timestamped
- language
- task annotations
operators ──drive──► robots ──capture──► teleop + ego data
▲ │
└────────── autonomy ◄── trained humanoids ──┘schema · ros2 / mcap · realsense calibration · sourced from HIW-500 (Unitree G1) · CC-BY-4.0
03SAMPLE EPISODES
Real captured episodes — three synchronized camera feeds with live joint, end-effector, and teleop-action telemetry. The raw signal, exactly as logged.
connecting feed…
04ACCESS
Source teleop data for your models.
Tell us the tasks and embodiments you're training.