can_to_martino_3

Whole-body teleoperation on a Unitree G1, recorded the evening of 2026-08-25. The robot picks a can off a low table and places it on a white table. Same rig, task and pipeline as nepyope/can_to_martino_2, recorded in three later sessions and released separately rather than merged into it.

episodes	29
frames	57,651
fps	50
codebase version	v3.0
task	Bring the can to the white table

Schema

Identical to can_to_martino_2, so the two can be concatenated.

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

As in can_to_martino_2, observation.state[29] and action[64] are identically 0.0 in every frame — these 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. The right gripper is healthy (closed in 40% of frames). No other channel is degenerate.

Provenance

Merged from three sessions. Episodes the operator discarded during collection were dropped, along with two stubs of 3.3 s and 0.1 s. The discard rate in these sessions was high — 49 of 80 recorded episodes, and 37 of 53 in the first session alone — so this data is a smaller, more heavily filtered sample than its recording time suggests.