AbdullahRasul/g1_live_tracking_restored

Rokoko → Unitree G1 + Dex3

This dataset was created using LeRobot from saved Rokoko suit/Smart Gloves teleoperation episodes in MuJoCo.

3 episodes · 3783 frames · 30 FPS · 2 cameras

Robot: Rokoko_G1_Dex3_FullBody. Dataset format: v3.0. Exporter runtime: LeRobot 0.4.1.

Tasks

• Transfer box from Table A to Table B

Observation and action semantics

• observation.state: Actual simulated/displayed kinematic G1 joint state at source snapshot; not hardware feedback.
• action: Associated published Rokoko-retargeted G1 target; not measured hardware command execution.
• Joint positions are radians. Root position is metres; root quaternion is (w, x, y, z).
• These are simulation demonstrations, not physical robot feedback or verified hardware trajectories.

Cameras

Feature	Resolution	Encoding	FPS
observation.images.front	480 × 360	av1, yuv420p	30
observation.images.side	480 × 360	av1, yuv420p	30

front is the robot-mounted camera; side is the external environment camera. Browser video decoding support depends on the browser and the listed codec.

Sampling

Timestamp-aware causal whole-snapshot hold, fixed FPS; no joint/pose/image interpolation; pause segments split. Original RGB JPEG snapshots encoded as video at export FPS; temporal upsampling does not create new images. Source timestamps and source frame IDs remain available as custom features; the complete resampling map is in provenance/resampling.jsonl.

Feature schema

Feature	Type	Shape
observation.state	float32	[43]
action	float32	[43]
observation.root_position	float32	[3]
observation.root_quaternion	float32	[4]
observation.target_root_position	float32	[3]
observation.target_root_quaternion	float32	[4]
observation.full_body_state	float32	[43]
observation.full_body_target	float32	[43]
observation.box_pose	float32	[7]
observation.grasp_state	float32	[3]
observation.glove_curl	float32	[2]
observation.solver	float32	[4]
observation.solver_valid	float32	[4]
observation.tracking	float32	[2]
source.timestamp	float64	[1]
source.capture_monotonic	float64	[1]
source.frame_index	int64	[1]
source.raw_index	int64	[1]
observation.human_positions	float32	[189]
observation.filtered_positions	float32	[189]
observation.human_rotations	float32	[567]
observation.filtered_rotations	float32	[567]
observation.images.front	video	[360, 480, 3]
observation.images.side	video	[360, 480, 3]
timestamp	float32	[1]
frame_index	int64	[1]
episode_index	int64	[1]
index	int64	[1]
task_index	int64	[1]

Complete metadata and dimension names: meta/info.json.

Joint ordering

Version: g1_29_dex3_14_mjcf_v1.

The same index ordering applies to observation.state and action.

Index	Feature name	Source joint	Units
0	left_hip_pitch_joint	left_hip_pitch_joint	radians
1	left_hip_roll_joint	left_hip_roll_joint	radians
2	left_hip_yaw_joint	left_hip_yaw_joint	radians
3	left_knee_joint	left_knee_joint	radians
4	left_ankle_pitch_joint	left_ankle_pitch_joint	radians
5	left_ankle_roll_joint	left_ankle_roll_joint	radians
6	right_hip_pitch_joint	right_hip_pitch_joint	radians
7	right_hip_roll_joint	right_hip_roll_joint	radians
8	right_hip_yaw_joint	right_hip_yaw_joint	radians
9	right_knee_joint	right_knee_joint	radians
10	right_ankle_pitch_joint	right_ankle_pitch_joint	radians
11	right_ankle_roll_joint	right_ankle_roll_joint	radians
12	waist_yaw_joint	waist_yaw_joint	radians
13	waist_roll_joint	waist_roll_joint	radians
14	waist_pitch_joint	waist_pitch_joint	radians
15	left_shoulder_pitch_joint	left_shoulder_pitch_joint	radians
16	left_shoulder_roll_joint	left_shoulder_roll_joint	radians
17	left_shoulder_yaw_joint	left_shoulder_yaw_joint	radians
18	left_elbow_joint	left_elbow_joint	radians
19	left_wrist_roll_joint	left_wrist_roll_joint	radians
20	left_wrist_pitch_joint	left_wrist_pitch_joint	radians
21	left_wrist_yaw_joint	left_wrist_yaw_joint	radians
22	left_hand_thumb_0_joint	left_hand_thumb_0_joint	radians
23	left_hand_thumb_1_joint	left_hand_thumb_1_joint	radians
24	left_hand_thumb_2_joint	left_hand_thumb_2_joint	radians
25	left_hand_middle_0_joint	left_hand_middle_0_joint	radians
26	left_hand_middle_1_joint	left_hand_middle_1_joint	radians
27	left_hand_index_0_joint	left_hand_index_0_joint	radians
28	left_hand_index_1_joint	left_hand_index_1_joint	radians
29	right_shoulder_pitch_joint	right_shoulder_pitch_joint	radians
30	right_shoulder_roll_joint	right_shoulder_roll_joint	radians
31	right_shoulder_yaw_joint	right_shoulder_yaw_joint	radians
32	right_elbow_joint	right_elbow_joint	radians
33	right_wrist_roll_joint	right_wrist_roll_joint	radians
34	right_wrist_pitch_joint	right_wrist_pitch_joint	radians
35	right_wrist_yaw_joint	right_wrist_yaw_joint	radians
36	right_hand_thumb_0_joint	right_hand_thumb_0_joint	radians
37	right_hand_thumb_1_joint	right_hand_thumb_1_joint	radians
38	right_hand_thumb_2_joint	right_hand_thumb_2_joint	radians
39	right_hand_middle_0_joint	right_hand_middle_0_joint	radians
40	right_hand_middle_1_joint	right_hand_middle_1_joint	radians
41	right_hand_index_0_joint	right_hand_index_0_joint	radians
42	right_hand_index_1_joint	right_hand_index_1_joint	radians

Machine-readable mapping: meta/robot_joint_mapping.json.

Dataset structure

README.md
meta/info.json
meta/stats.json
meta/tasks.parquet
meta/episodes/chunk-*/file-*.parquet
data/chunk-*/file-*.parquet
videos/observation.images.{front,side}/chunk-*/file-*.mp4
provenance/export.json
provenance/resampling.jsonl
validation.json

Human skeleton/glove data, root transforms, box/grasp/task state and solver diagnostics are additional features. Source episode status, session references and calibration metadata are retained in provenance. Missing solver metrics use a validity mask.

Limitations and source quality

Acceptance denotes a collection decision, not a guarantee of task success or robot safety. Temporal upsampling does not add camera observations. Collision diagnostics describe the source solver; they do not certify physical robot execution.

• accepted/episode_000000_attempt_000: first-frame object/grasp/success state is not reset; source preserved as collected.

Visualization

Use the official LeRobot visualizer and enter the uploaded dataset repository ID. All files must be uploaded at the dataset repository root.

License

Not specified. The dataset owner must choose a license before licensed redistribution.