THULab/Jiangeng_G1_06_11

G1 06 11 TsFile

Apache TsFile edition of Jiangeng/G1_06_11, a LeRobot v2.1 Unitree G1 whole-body dataset for the task: Put the teddy bear in the bag. Numeric trajectories are stored in one table-model TsFile.

Source and attribution

• Original repository owner, publisher, uploader, and sole listed contributor: Jiangeng. No separate personal name is provided.
• License: the source repository does not declare one.
• Paper, homepage, and citation: the source repository does not provide them.
• Split: train; 51 episodes; 79,948 frame rows; one task; 50 Hz; 51 source Parquet episode shards.
• Task 0: Put the teddy bear in the bag.

Data layout

The table is jiangeng_g1_06_11 and contains 79,948 rows across 51 TAG devices. The source Parquet shards total 67,782,033 bytes; the TsFile is 31,817,362 bytes (46.9% of the source Parquet size).

Column or group	TsFile role	Type	Meaning
Time	TIME	INT64 milliseconds	round(timestamp * 1000), restarting at zero for each episode
episode_index	TAG	STRING from source INT64	Source episode identity, values 0 through 50
task_index	TAG	STRING from source INT64	Source task identity, value 0
frame_index	FIELD	INT64	Frame position within the episode
sample_index	FIELD	INT64	Source index, renamed for clarity
teleop_delta_heading	FIELD	DOUBLE	Source heading delta
teleop_smpl_frame_index	FIELD	INT64	Source SMPL frame index
teleop_stream_mode, teleop_planner_mode	FIELD	INT32	Teleoperation and locomotion planner modes
teleop_planner_speed, teleop_planner_height	FIELD	FLOAT	Planner speed and height
observation_state_0 ... observation_state_42	FIELD	FLOAT	G1 joint state; 43 elements
observation_eef_state_0 ... observation_eef_state_13	FIELD	FLOAT	left/right wrist position and quaternion state; 14 elements
action_wbc_0 ... action_wbc_42	FIELD	FLOAT	whole-body controller action; 43 elements
observation_root_orientation_0 ... observation_root_orientation_3	FIELD	FLOAT	flattened source observation.root_orientation vector; 4 elements
observation_projected_gravity_0 ... observation_projected_gravity_2	FIELD	FLOAT	flattened source observation.projected_gravity vector; 3 elements
observation_cpp_rotation_offset_0 ... observation_cpp_rotation_offset_3	FIELD	FLOAT	flattened source observation.cpp_rotation_offset vector; 4 elements
observation_init_base_quat_0 ... observation_init_base_quat_3	FIELD	FLOAT	flattened source observation.init_base_quat vector; 4 elements
action_motion_token_0 ... action_motion_token_63	FIELD	FLOAT	motion token; 64 elements
teleop_smpl_joints_0 ... teleop_smpl_joints_71	FIELD	FLOAT	SMPL joint values; 72 elements
teleop_smpl_pose_0 ... teleop_smpl_pose_62	FIELD	FLOAT	SMPL pose values; 63 elements
teleop_body_quat_w_0 ... teleop_body_quat_w_3	FIELD	FLOAT	flattened source teleop.body_quat_w vector; 4 elements
teleop_target_body_orientation_0 ... teleop_target_body_orientation_5	FIELD	FLOAT	flattened source teleop.target_body_orientation vector; 6 elements
teleop_left_hand_joints_0 ... teleop_left_hand_joints_6	FIELD	FLOAT	flattened source teleop.left_hand_joints vector; 7 elements
teleop_right_hand_joints_0 ... teleop_right_hand_joints_6	FIELD	FLOAT	flattened source teleop.right_hand_joints vector; 7 elements
teleop_left_wrist_joints_0 ... teleop_left_wrist_joints_2	FIELD	FLOAT	flattened source teleop.left_wrist_joints vector; 3 elements
teleop_right_wrist_joints_0 ... teleop_right_wrist_joints_2	FIELD	FLOAT	flattened source teleop.right_wrist_joints vector; 3 elements
teleop_planner_movement_0 ... teleop_planner_movement_2	FIELD	FLOAT	flattened source teleop.planner_movement vector; 3 elements
teleop_planner_facing_0 ... teleop_planner_facing_2	FIELD	FLOAT	flattened source teleop.planner_facing vector; 3 elements
teleop_vr_3pt_position_0 ... teleop_vr_3pt_position_8	FIELD	FLOAT	flattened source teleop.vr_3pt_position vector; 9 elements
teleop_vr_3pt_orientation_0 ... teleop_vr_3pt_orientation_17	FIELD	FLOAT	three-point VR orientation; 18 elements

The source timestamp column is not stored separately because it is represented by Time / 1000 seconds. All 79,948 numeric rows, every episode and task identity, frame_index, and every numeric vector element are retained. Source vector prefixes are preserved with dots replaced by underscores.

Encoding and compression

• TIME, INT32, and INT64: TS_2DIFF + LZ4
• FLOAT and DOUBLE: GORILLA + LZ4
• BOOLEAN, when present: RLE + LZ4
• TAG values: TsFile table-model device/tag storage

Videos and alignment

The 51 source ego-view MP4 files are not included here. They remain in the original repository at videos/chunk-000/observation.images.ego_view, named episode_000000.mp4 through episode_000050.mp4. The source directory is displayed as 257 MB on Hugging Face.

The source template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index and frame_index to align a numeric row with its 50 FPS video frame.

Read example

from tsfile import TsFileReader

reader = TsFileReader("data/jiangeng_g1_06_11.tsfile")
with reader.query_table(
    "jiangeng_g1_06_11",
    [
        "episode_index",
        "task_index",
        "frame_index",
        "sample_index",
        "observation_state_0",
        "action_wbc_0",
    ],
    batch_size=1024,
) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()