can_to_martino_2

Whole-body teleoperation on a Unitree G1, recorded 2026-08-25. The robot picks a can off a low table and places it on a white table. Successor to nepyope/can_to_martino (44 episodes), recorded on the same rig with the same pipeline.

episodes	95
frames	187,224
fps	50
codebase version	v3.0
task	Bring the can to the white table

Schema

feature	dtype	shape	contents
observation.state	float32	[31]	29 body joints in G1_29_JointIndex order, then left/right gripper
action	float32	[66]	64-D SONIC motion token, then left/right gripper
observation.images.ego_view	video	[480, 640, 3]	head camera
observation.images.left_wrist	video	[480, 640, 3]	left wrist camera
observation.images.right_wrist	video	[480, 640, 3]	right wrist camera

action[t] is the command that produces observation.state[t+1], so each episode is one frame shorter than its recording.

Known issue: the left gripper is dead

observation.state[29] and action[64] are identically 0.0 in every frame — the episodes were effectively recorded one-handed. Their q01/q99 in meta/stats.json are set by hand to 0/1 so quantile normalisation does not divide by zero. A policy trained on this data will never open or close the left hand. The right gripper is healthy (closed in 38% of frames).

No other channel is degenerate: all 64 token dimensions and all 29 joints vary.

Provenance

Merged from 11 recording sessions. Episodes the operator discarded during collection were dropped, along with one 2.2-second stub. The sessions used five different task prompts for what is the same behaviour; all episodes were relabelled to a single prompt.