Datasetpublic

lewislf/G1_Dex3_pick_white_cup_right8_1squeeze

G1_Dex3_pick_white_cup_right8_1squeeze Teleoperated demonstration dataset for training a Vision-Language-Action (VLA) policy on the Unitree G1 humanoid robot with a Dex3 dexterous hand. The task is to pick up a white cup from a table using only the right arm. Task Pick up the white cup The robot reaches toward a white cup placed on a table in front of it, grasps it with the right hand (index + middle fingers closing against the thumb), and lifts it.… See the full description on the dataset page: https://huggingface.co/datasets/lewislf/G1_Dex3_pick_white_cup_right8_1squeeze.

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes75downloads

Dataset Card

G1Dex3pickwhitecupright81squeeze

Teleoperated demonstration dataset for training a Vision-Language-Action (VLA) policy on the Unitree G1 humanoid robot with a Dex3 dexterous hand. The task is to pick up a white cup from a table using only the right arm.

Task

Pick up the white cup

The robot reaches toward a white cup placed on a table in front of it, grasps it with the right hand (index + middle fingers closing against the thumb), and lifts it.

Dataset at a glance

Property	Value
Robot	Unitree G1 + Dex3 right hand
Episodes	238
Total frames	52,952
FPS	30
Avg episode duration	~7.4 s (median 6.9 s)
Episode duration range	3.9 s – 26.5 s
LeRobot codebase version	v3.0
Data size	~100 MB parquet + ~200 MB video

Action space (8 dims)

The policy outputs joint position targets for the right arm and a single grasp scalar:

Dim	Name	Description
0	`kRightShoulderPitch.q`	Right shoulder pitch
1	`kRightShoulderRoll.q`	Right shoulder roll
2	`kRightShoulderYaw.q`	Right shoulder yaw
3	`kRightElbow.q`	Right elbow
4	`kRightWristRoll.q`	Right wrist roll
5	`kRightWristPitch.q`	Right wrist pitch
6	`kRightWristYaw.q`	Right wrist yaw
7	`right_grasp_squeeze.q`	Grasp squeeze (0 = open, 1 = fully closed)

All arm joints in radians. The squeeze scalar drives index and middle fingers via a fixed mapping; thumb position is mechanically coupled.

Observation space

`observation.state` (14 dims)

Proprioceptive state: the first 7 dims mirror the action arm joints (measured positions), plus 7 finger joint angles (thumb × 3, index × 2, middle × 2).

`observation.images.head_camera`

RGB video stream from the head camera. Resolution: 480 × 848 px, codec H.264/YUV420p, 30 fps.

`observation.images.head_camera_depth`

Aligned depth map (480 × 848 × 1), stored as PNG per frame. Units: millimeters (uint16).

`observation.left_hand_pressure` / `observation.right_hand_pressure`

Raw readings from the Dex3 tactile sensor array: 108 sensors per hand (float32). Right-hand pressure shows contact activity during the grasp phase.

Action statistics (global)

Dim	Mean	Std	Min	Max
ShoulderPitch	-0.135	0.308	-0.908	0.682
ShoulderRoll	-0.109	0.128	-0.751	0.364
ShoulderYaw	-0.072	0.189	-0.961	0.431
Elbow	0.121	0.597	-0.956	1.330
WristRoll	-0.002	0.126	-0.440	0.595
WristPitch	-0.355	0.317	-1.051	0.655
WristYaw	0.046	0.188	-0.711	0.875
GraspSqueeze	0.411	0.487	0.0	1.0

Data collection

Data was collected via VR teleoperation (Quest 2 + custom XR bridge) at 30 Hz. Each episode starts with the arm at a neutral pose, the robot reaches for and picks up the cup, and the episode ends after the cup is lifted.

LeRobot compatibility

This dataset follows the LeRobot v3.0 data format and can be loaded directly:

python

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("lewislf/G1_Dex3_pick_white_cup_right8_1squeeze")

License

MIT