Datasetpublic

tysyuvraj/g1-apple2plate-mixed-v3

g1-apple2plate-mixed-v3 A multi-task teleoperation dataset on a Unitree G1 humanoid covering apple↔plate/table manipulation, target corrections, and task cancellation. Recorded for fine-tuning NVIDIA GR00T N1.7. Stored in LeRobot v2.1 format (what GR00T's loader consumes). Robot Embodiment: Unitree_G1_Dex1_UpperBody_2Cam (Dex1 parallel grippers) State / action: 16-D — left arm 0:7, right arm 7:14, left gripper 14:15, right gripper 15:16 Cameras: cam_head… See the full description on the dataset page: https://huggingface.co/datasets/tysyuvraj/g1-apple2plate-mixed-v3.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes327downloads

Dataset Card

g1-apple2plate-mixed-v3

A multi-task teleoperation dataset on a Unitree G1 humanoid covering apple↔plate/table manipulation, target corrections, and task cancellation. Recorded for fine-tuning NVIDIA GR00T N1.7. Stored in LeRobot v2.1 format (what GR00T's loader consumes).

Robot

—Embodiment: Unitree_G1_Dex1_UpperBody_2Cam (Dex1 parallel grippers)
—State / action: 16-D — left arm 0:7, right arm 7:14, left gripper 14:15, right gripper 15:16
—Cameras: cam_head, cam_right_wrist @ 30 Hz

—477 episodes, 65,502 frames, 30 fps, LeRobot codebase v2.1

Tasks (7 distinct language instructions)

Instruction	Source behaviors	Episodes
`Pick up the red apple and place it on the plate.`	red pick + pre/post-grasp corrections ending red	~100
`Pick up the green apple and place it on the plate.`	green pick + pre/post-grasp corrections ending green	~101
`Pick up both apples and place them on the plate.`	sequential place	43
`Pick up the red apple from the plate and place it on the table.`	return	51
`Pick up the green apple from the plate and place it on the table.`	return	61
`Pick up both apples from the plate and place them on the table.`	sequential return	39
`Stop the current task and return to the home position.`	cancel (red or green)	82

Correction & cancel design

—Pre-grasp correction (arm redirects toward the other apple before grasping) and post-grasp correction (switch target while already holding one apple) are both labeled with the final intended target's normal instruction — not a special "correction" phrase. At deployment the controller simply sends the new target's task string; visual context disambiguates. This trains the policy to recover to the correct object under one instruction.
—Cancel episodes (abort after starting either apple) share a single instruction, since the desired behavior — stop and return home — is identical regardless of which apple was involved.

Usage

python

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("tysyuvraj/g1-apple2plate-mixed-v3")
print(ds.meta.total_episodes, ds.meta.total_frames)

Requires a LeRobot version that reads codebase v2.1 (e.g. lerobot==0.3.3). meta/modality.json maps the 16-D state/action into GR00T's arm/gripper groups and the two cameras, so it can be passed directly to GR00T N1.7 fine-tuning with embodiment tag new_embodiment.

Notes