4in12

Unitree G1 (29-DoF humanoid + Dex3 hands) whole-body teleoperation demonstrations, recorded with MA-Teleop: a human in a PICO VR headset drives the robot through the GR00T-WholeBodyControl / SONIC whole-body policy; the robot records synchronized camera frames and proprioception, the PC records the motion tokens that drove it. LeRobot v3.0 dataset.

Episodes	20
Frames	10758 @ 10 Hz
Robot	g1 (Unitree G1 EDU, 29 DoF, Dex3 hands)
Cameras	observation.images.egocentric 720×1280 (Stereolabs ZED head camera over UVC, undistorted)
Task instruction (as stored)	"4in2"

Features

• states [43] — 29 measured body joint positions (MuJoCo/motor order, rad) + left hand 7 + right hand 7 joints (commanded).
• action [78] — the 64-d SONIC motion token in effect at the frame (nearest-past token, zero-order hold) + left/right hand 7+7 joints.
• observation.images.egocentric — head camera video (AV1, in videos/).
• meta/modality.json — GR00T-style slicing (state 0–43, action 0–78, video ego_view).

Sessions

Several recording sessions of the same task are appended in chronological order with contiguous episode numbering:

session ({task}_{YYYY-MM-DD_HH-MM-SS})	episodes	count
4in12_2026-10-12_13-23-04	episode_0 – episode_19	20

Provenance

Robot half (data_episode_N.pkl + video_episode_N.mp4, Jetson) and PC half (tokens_episode_N.npz, absolute-epoch timestamps) are aligned per frame by timestamp (data_pipeline/g1/align_episode.py, PC and robot clocks chrony-synced), appended with prepare_task.py and built with build_lerobot_v3.py.