Evanxw0715/BiCAT-UnitreeG1-Dataset

BiCAT Unitree G1 Dataset

Bimanual Unitree G1 Dex1 bottle- and plate-manipulation demonstrations in LeRobot v3.0 format. Each task folder is a complete dataset with its own meta/, data/, and videos/ directories. Episode indices start at 0 within each task.

Folder	Task	Episodes	Frames	FPS	Size
bottles/	Bottle manipulation	100	44,123	30	0.686 GB
plate/	Place the bowl on the plate.	100	53,822	30	1.427 GB

Total dataset size: approximately 2.113 GB, counting files under bottles/ and plate/. GB uses decimal units (1 GB = 1,000,000,000 bytes).

Path	Contents
README.md	Dataset card and loading instructions
bottles/data/	Frame-level observations, actions, timestamps, and indices
bottles/meta/	Dataset information, tasks, statistics, episode references, and source provenance
bottles/videos/	Top, left-wrist, and right-wrist camera videos
plate/data/	All 53,822 frame rows in one Parquet file
plate/meta/	Instruction, schema, statistics, episode references and validation reports
plate/videos/	Three concatenated AV1 videos, one per camera

The dataset uses the unitree_g1_dex1 robot. Its main features are:

Feature	Shape	Description
observation.images.left, observation.images.right	480 × 640 × 3	RGB video, 30 FPS
observation.images.top	360 × 640 × 3	RGB video, 30 FPS
observation.state_joint_angles, action_joint_angles	16	Seven joints and one gripper per arm
observation.state_eef_absolute, action_eef_absolute, action_eef_delta	16	Position, quaternion in qw, qx, qy, qz order, and gripper per arm
observation.velocity, observation.effort	16	Joint and gripper velocity/effort observations
observation.full_body_q, observation.full_body_dq, observation.full_body_tau_est	35	Full-body position, velocity, and estimated torque telemetry

Additional timing, camera sequence, robot tick, and episode outcome fields are retained. Exact feature names, dtypes, dimensions, and component order are defined in each task folder’s meta/info.json.

The 100 bottle episodes retain the nine source task labels: Bottle 01, Bottle 02, Bottle 03, Bottle 04, Bottle 05, Bottle 06, Bottle 07, Bottle 09, and Bottle 10. These labels identify the source tasks; they have not been replaced with a new natural-language instruction. Episode indices run from 0 to 99. Source mappings and per-episode metadata are retained under bottles/meta/.

During initial preparation, frame-level data was compacted into bottles/data/chunk-000/file-000.parquet. Each camera view has one AV1 video at bottles/videos/{video_key}/chunk-000/file-000.mp4. Episode data-file references and video-file references were updated together. All 44,123 frame rows and all 100 episode slices were compared against the original data; the videos were copied without re-encoding and verified by SHA256. Results are recorded in bottles/meta/compaction_report.json. Earlier merge and video reports document preparation of the source dataset before compaction.

G1 and YAM share the LeRobot v3.0 storage layout and the main end-effector feature names. G1 has 16 joint-action dimensions, whereas the YAM dataset has 14; use the G1 feature definitions when configuring a policy. The retained G1 kinematics metadata describes a virtual end-effector frame rather than a calibrated fingertip/TCP.

Load a task locally

The LeRobot loader expects a dataset root containing meta/info.json. Point it at one task folder after downloading the repository:

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

snapshot = Path(snapshot_download(
    repo_id="Evanxw0715/BiCAT-UnitreeG1-Dataset",
    repo_type="dataset",
))
bottles = LeRobotDataset("local/bottles", root=snapshot / "bottles")
plate = LeRobotDataset("local/plate", root=snapshot / "plate")
assert plate[0]["task"] == "Place the bowl on the plate."

The videos are grouped into files; episode video locations and timestamps are recorded in meta/episodes/. Use a LeRobot v3.0-compatible loader and an AV1-capable video backend.

Bottle video timestamp correction — 2026-10-22

The three concatenated AV1 videos have been losslessly remuxed with exact global timestamps frame_index / 30. Compressed image packets are unchanged; the existing compacted data Parquet, episode metadata, task labels, paths and episode offsets are preserved.

All 44,123 frame timestamps per camera pass the original 0.0001-second tolerance. Maximum float32 comparison error is 0.0000152587890625 seconds. The original LeRobot 0.4.4 reader, including TorchCodec approximate seek and PyAV, passed 301 samples across all 100 episodes and the previously failing positions; 903 camera frames per backend produced identical pixels. No reader patch or tolerance increase is required for the validated reader version.

See bottles/meta/concat_cfr_timestamp_audit.json, bottles/meta/concat_cfr_reader_verification.json and bottles/meta/video_remux_report.json. Historical compaction reports describe the previous preparation step. Download the latest repository revision into a fresh snapshot/root to avoid reusing old video files. This validates dataset reading, not a completed server training run.

Plate dataset

plate/ contains 100 episodes from plate_01 through plate_07, after excluding plate_01 ordinals 1, 12 and 13 and plate_03 ordinal 5 (one-based numbering). The shared instruction is Place the bowl on the plate. It is stored in plate/meta/tasks.parquet; every frame uses task_index = 0, and all episode task labels match. Source collection labels are retained in the provenance mapping.

The data is stored in a single plate/data/chunk-000/file-000.parquet. Each camera has one losslessly concatenated AV1 video with exact CFR timestamps. All 53,822 frame timestamps per view pass the original tolerance. After instruction assignment, the original LeRobot reader passed 300 samples (900 images per backend), with identical PyAV/TorchCodec pixels. See the plate README for the file layout, source counts, loading example and validation reports.