zjt24/gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5

GR00T G1 Grab-Bottle Right Hand Zero-Wandering Radius-5 TsFile

This dataset is a TsFile conversion of the Hugging Face dataset cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-5. It contains the numeric frame-level time-series data from the LeRobot v2.1 Unitree G1 right-hand bottle-grab task. The source dataset is the radius-5 zero-wandering curation of a 105-episode teleoperation set, where each surviving clean sub-segment is emitted as its own episode.

Modalities: Time-series. The original dataset also contains video, but videos are not included in this TsFile conversion.

Source Dataset

• Original dataset: cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-5
• Source format: LeRobot v2.1
• License: Apache-2.0
• Task: "grab the bottle" with the right hand
• Robot and setup: Unitree G1 with Inspire FTP hands, SONIC whole-body controller, PICO 4 Ultra teleoperation, and a head-mounted Intel RealSense D435I camera
• Source scale: 417 clean-segment episodes, 48,577 frames, 1 task, 50 fps
• Source curation: zero-wandering smoothing half-width 5 frames, safe_frames = 40
• Source video stream: observation.images.ego_view, 640 x 480 H.264 at 50 fps

Converted Files

• data/gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5_train.tsfile
• Table name: gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5_train
• Split: train, from the source split specification 0:417
• TsFile size after conversion: 41,918,831 bytes
• Converted rows: 48,577
• Converted schema width: 388 columns total, including 1 Time column, 2 TAG columns, and 385 FIELD columns

The source meta/ directory is mirrored in this repository. meta/info.json has been updated so data_path points to the converted TsFile artifact and tsfile_conversion documents the mapping from the original LeRobot data.

Schema

• Time: integer milliseconds, computed as round(timestamp * 1000). Time restarts within each episode, matching the source per-episode timeline.
• TAG columns: episode_index, task_index.
• FIELD columns: scalar frame and teleoperation measurements plus flattened robot observation/action vectors.
• frame_index is retained as a FIELD column.
• Source column index is renamed to sample_index.
• Source column timestamp is dropped after creating Time, because it is redundant with Time / 1000 seconds.

Vector columns are flattened into scalar TsFile fields by preserving the full source column name, replacing . with _, and appending the element index. Examples:

observation.state[43]        -> observation_state_0 ... observation_state_42
observation.eef_state[14]    -> observation_eef_state_0 ... observation_eef_state_13
action.wbc[43]               -> action_wbc_0 ... action_wbc_42
action.motion_token[64]      -> action_motion_token_0 ... action_motion_token_63
teleop.* vectors             -> teleop_*_0 ... teleop_*_N

Additional flattened vector groups include observation.root_orientation[4], observation.projected_gravity[3], observation.cpp_rotation_offset[4], observation.init_base_quat[4], teleop.smpl_joints[72], teleop.smpl_pose[63], teleop.body_quat_w[4], teleop.target_body_orientation[6], teleop.left_hand_joints[7], teleop.right_hand_joints[7], teleop.left_wrist_joints[3], teleop.right_wrist_joints[3], teleop.planner_movement[3], teleop.planner_facing[3], teleop.vr_3pt_position[9], and teleop.vr_3pt_orientation[18].

Video Policy

Videos are intentionally not uploaded here. The source video feature observation.images.ego_view remains available in the original dataset under videos/. The TsFile artifact stores the aligned numeric state/action/time/task data.

Validation

Conversion was validated locally with the project pipeline:

.\.venv\Scripts\python.exe scripts\pipeline.py inspect configs\gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5.yaml
.\.venv\Scripts\python.exe scripts\pipeline.py convert configs\gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5.yaml
.\.venv\Scripts\python.exe scripts\pipeline.py validate configs\gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5.yaml

The readback check opened the TsFile successfully, confirmed 417 devices, and matched the TsFile metadata row count of 48,577 rows against the staged Parquet row count.

Read Example

from tsfile import TsFileReader

path = "data/gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5_train.tsfile"
table = "gr00t_g1_grab_bottle_right_hand_zero_wandering_smooth_radius_5_train"

with TsFileReader(path) as reader:
    columns = [
        "episode_index",
        "task_index",
        "frame_index",
        "sample_index",
        "action_motion_token_0",
        "observation_state_0",
    ]
    with reader.query_table(table, columns, batch_size=4096) as result:
        batch = result.read_arrow_batch()
        print(batch)

Citation

The source dataset card includes BibTeX entries for this dataset, GR00T N1, SONIC, and LeRobot. Please cite the original dataset when using this conversion.