THULab/unitreerobotics_G1_Dex3_BlockStacking_Dataset

G1 Dex3 Block Stacking TsFile

This repository is an Apache TsFile conversion of `unitreerobotics/G1_Dex3_BlockStacking_Dataset`, a LeRobot manipulation dataset for stacking three cubic blocks on a desktop in red, yellow, and blue order.

Source dataset

• —Original dataset: https://huggingface.co/datasets/unitreerobotics/G1Dex3BlockStacking_Dataset
• —Author / repository owner: Unitree Robotics (unitreerobotics)
• —Source-tree contributors shown by Hugging Face: Henry-Ellis, lv-1
• —License: Apache-2.0
• —Robot: Unitree G1, 7-DOF dual-arm, with three-finger Dex3 hands
• —Recording frequency: 30 Hz
• —Cameras: two head-mounted streams (cam_left_high, cam_right_high) and two wrist-mounted streams (cam_left_wrist, cam_right_wrist)
• —Resolution: 640x480 RGB; source video codec is AV1
• —Split: train (0:301 episodes)
• —Episodes / rows / tasks: 301 / 281,196 / 1
• —Source data shards: 1 chunk / 1 Parquet file; source card reports 1,204 video files
• —Source frame file: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet

The task description in the original card is to stack three 5 cm cubic blocks on black tape on the desktop from bottom to top in red, yellow, and blue order.

Converted files

• —TsFile: data/unitreerobotics_g1_dex3_blockstacking.tsfile
• —Table: unitreerobotics_g1_dex3_blockstacking
• —Rows: 281,196
• —Episodes: 301
• —Episode files: 1 merged table-model TsFile
• —Time precision: milliseconds
• —Source metadata: meta/tasks.parquet, meta/episodes/..., and meta/stats.json are retained; meta/info.json documents the conversion.

Schema

Time = round(timestamp * 1000). The source timestamp is in seconds and restarts at zero for each episode, so Time is monotonic within each (episode_index, task_index) device. The source timestamp column is dropped after synthesis because it is redundant with Time / 1000.

Vector columns are flattened row-major while preserving source prefixes:

• —observation.state[28] -> observation_state_0 ... observation_state_27
• —action[28] -> action_0 ... action_27

The source index is renamed to sample_index; frame_index, episode_index, and task_index are preserved. No numeric state/action rows are dropped. Image/video feature columns are not written into TsFile because the local source Parquet contains numeric references only.

Encoding and compression

The compact TsFile profile is used to avoid producing a file larger than the source Parquet:

• —FLOAT/DOUBLE: GORILLA + LZ4
• —INT32/INT64: TS_2DIFF + LZ4
• —Time: TS_2DIFF + LZ4
• —BOOLEAN: RLE + LZ4 (no BOOLEAN fields are present in this dataset)
• —TAG columns: TsFile table/device TAG mechanism

The staged Parquet and local validation report are kept outside the upload set under .conversion_work/.

Videos and attribution

Videos are not included in this TsFile repository. They remain in the original Hugging Face dataset at:

• —`videos/observation.images.cam_left_high`
• —`videos/observation.images.cam_left_wrist`
• —`videos/observation.images.cam_right_high`
• —`videos/observation.images.cam_right_wrist`

The source metadata video pattern is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. episode_index and frame_index remain unchanged in the TsFile so consumers can align a time-series row with its original video frame.

Read example

import sys
sys.path.insert(0, r"D:\\code\\.deps")
from tsfile import TsFileReader

path = r"D:\\code\\py\\unitreerobotics_G1_Dex3_BlockStacking_Dataset\\data\\unitreerobotics_g1_dex3_blockstacking.tsfile"
reader = TsFileReader(path)
table = reader.get_all_table_schemas()["unitreerobotics_g1_dex3_blockstacking"]
columns = [c.get_column_name() for c in table.get_columns() if c.get_column_name() != "Time"]
with reader.query_table("unitreerobotics_g1_dex3_blockstacking", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()

Citation

Please cite the original Unitree Robotics dataset repository and the LeRobot project when using this conversion. No separate paper or BibTeX entry was provided in the source dataset card.