jren313/g1-pipette-tip-teleop

G1 pipette teleop — LeRobot v2.1

64 episodes, 56742 frames at 60 fps (15.8 min), robot unitree_g1_29dof_inspire.

Converted 2026-08-04T17:07:55Z from /media/jren313/Crucial X10/08_03_2026_Robot_Data (sessions: 20260803_1445, 20260803_1455, 20260803_1511) by scripts/convert_g1_recording_to_lerobot.py in VLAPolicyBridge.

Task

• goal — Attach a red pipette tip from the tip rack to the pipette, then eject the used tip into the tip disposal box.
• desc — A humanoid robot operates a robotic hand to attach and eject the red tip on a pipette.
• steps — Step 1: Align the pipette with the red tip in the tip rack. Step 2: Apply force to attach the tip securely to the pipette. Step 3: Lift the pipette tip out of the tip well. Step 4: Press the tip ejector to discard the tip into the disposal box on the side.

All three are rows in meta/tasks.jsonl, reachable per frame as the annotation.human.task_description / .desc / .steps columns. task_index is 0 (the goal), so anything that reads only the standard LeRobot task string gets the goal.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("jren313/g1-pipette-tip-teleop")
frame = ds[0]
frame["observation.images.rgb"]   # (3, 720, 1280) float32
frame["observation.state"]        # (34,) 29 joints + imu_ang_vel + imu_rp
frame["action"]                   # (35,) 29 joint targets + root cmd
frame["action.hand_right"]        # (6,) Inspire, raw 0-1000 register units

Hand channels are [pinky, ring, middle, index, thumb_bend, thumb_rot] in raw Inspire RH56 register units (0-1000), not radians.

Depth is not a LeRobot feature (see below); read it directly:

import numpy as np
with np.load('depth/chunk-000/episode_000000.npz') as z:
    depth = z['000123']    # (720, 1280) uint16, same instant as frame_index 123

Layout

meta/info.json              LeRobot v2.1 descriptor (features, fps, paths)
meta/tasks.jsonl            goal / desc / steps
meta/episodes.jsonl         per-episode length + task
meta/episodes_stats.jsonl   per-episode stats (v2.1)
meta/stats.json             aggregated stats (v2.0 consumers)
meta/modality.json          GR00T's view of the columns
meta/source_schema.json     the recorder's original schema, verbatim
meta/source_provenance.jsonl  new episode index -> source session/episode
meta/source_notes.json      conversion settings + caveats
data/chunk-000/episode_NNNNNN.parquet
videos/chunk-000/observation.images.rgb/episode_NNNNNN.mp4
depth/chunk-000/episode_NNNNNN.npz         16-bit lossless (see meta/depth_info.json)

Features

key	dtype	shape
observation.images.rgb	video	[720, 1280, 3]
observation.state	float32	[34]
observation.state.hand_left	float32	[6]
observation.state.hand_right	float32	[6]
observation.state.neck	float32	[2]
action	float32	[35]
action.hand_left	float32	[6]
action.hand_right	float32	[6]
action.neck	float32	[2]
action.low_level	float32	[29]
observation.effort	float32	[29]
observation.force.hand_left	float32	[6]
observation.force.hand_right	float32	[6]
observation.tactile.hand_left	int32	[1062]
observation.tactile.hand_right	int32	[1062]
observation.t_img_ms	int64	[1]
observation.t_action_ms	int64	[1]
annotation.human.task_description	int64	[1]
annotation.human.desc	int64	[1]
annotation.human.steps	int64	[1]
timestamp	float32	[1]
frame_index	int64	[1]
episode_index	int64	[1]
index	int64	[1]
task_index	int64	[1]

Caveats

Carried over from the recording as-is — these are properties of the capture, not of the conversion:

Channels that never change value (recorded, but carrying no signal). The most consequential is observation.state.hand_right: the right hand is the one doing the task, and its commanded action varies normally, but its joint feedback is frozen — so there is no right-hand proprioception to learn from.

channel	constant in
action.hand_left	64/64 episodes
observation.force.hand_left	63/64 episodes
observation.force.hand_right	64/64 episodes
observation.state.hand_left	37/64 episodes
observation.state.hand_right	64/64 episodes
observation.state.neck	64/64 episodes

• timestamp is derived from frame_index/fps; the recorder's own per-frame clocks are kept as observation.t_img_ms / observation.t_action_ms.
• The recorder's t_state was frozen for every episode and is not carried as a column; its constant value is in source_provenance.jsonl.
• Episodes were renumbered across sessions; source_provenance.jsonl maps back.