azithromycin/unitreerobotics_G1_Dex1_Clean_Table

G1 Dex1 Clean Table — Apache TsFile conversion

This repository is a compact Apache TsFile representation of the unitreerobotics/G1_Dex1_Clean_Table LeRobot dataset. It contains the robot state/action time series and episode/task metadata. The source camera videos are intentionally not copied here.

Source dataset and attribution

• Publisher/authors: Unitree Robotics; dataset contributors shown by Hugging Face are wangcong and wangcong627.
• License: Apache-2.0.
• Homepage: UnifoLM-VLA-0.
• Task: organize and tidy items on a table (7-DOF dual-arm G1, gripper end effectors).
• Acquisition: 30 Hz; 640×480 images; approximately 20–40 seconds per operation.
• Citation/paper: the original dataset card does not provide a BibTeX citation or paper reference.

The source has one train split with 200 episodes, 265,701 rows and one task. There are 200 source Parquet episode files in data/chunk-000/ and 800 source video files (four streams × 200 episodes). The original video layout is:

videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

For this dataset, video_key is one of observation.images.cam_left_high, cam_right_high, cam_left_wrist, or cam_right_wrist. See the original videos directory. Videos remain in the original Hugging Face dataset and are not included in this TsFile repository; align them to rows with episode_index and frame_index.

Converted artifact

• TsFile: data/unitreerobotics_G1_Dex1_Clean_Table.tsfile
• Table: unitreerobotics_G1_Dex1_Clean_Table (the Python SDK exposes the normalized lower-case table name)
• Rows: 265,701
• Episodes/devices: 200
• Time precision: integer milliseconds
• Source Parquet shards: 200 episode files, merged into one TsFile

Schema

Category	Columns
TIME	Time (INT64, ms)
TAG/device	episode_index (INT64), task_index (INT64)
FIELD scalars	frame_index, sample_index (INT64); four gripper scalars (FLOAT)
FIELD vectors	observation.left_arm/right_arm (7 each), observation.left_ee/right_ee (6 each), observation.body (29), and matching action.* vectors, flattened to scalar *_0 … *_{N-1} FLOAT fields

Dots in source names are replaced by underscores while preserving the full prefix (for example, observation.left_arm → observation_left_arm_0..observation_left_arm_6). The source index is renamed to sample_index.

Conversion details

• Time = round(timestamp * 1000) with millisecond precision. timestamp is dropped because it is redundant (Time / 1000 seconds); frame_index is kept.
• Rows are sorted by episode_index, task_index, then Time; Time is monotonic within each episode and restarts from zero at the first frame.
• episode_index and task_index are stored as TsFile TAG/device dimensions, not duplicated as ordinary fields.
• Numeric codec profile: FLOAT/DOUBLE → GORILLA, INT32/INT64 and Time → TS_2DIFF, all with LZ4 compression. No BOOLEAN source fields exist in this dataset; the configured BOOLEAN policy is RLE + LZ4.
• Dropped/omitted source data: only redundant timestamp is dropped from the tabular rows; four video columns are omitted from TsFile and remain at the source URL above. No numeric rows or measurements are intentionally removed.

Reading

from tsfile import TsFileReader

reader = TsFileReader("data/unitreerobotics_G1_Dex1_Clean_Table.tsfile")
table = next(iter(reader.get_all_table_schemas()))
columns = [c.get_column_name() for c in reader.get_all_table_schemas()[table].get_columns()
           if c.get_column_name() != "Time"]
with reader.query_table(table, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()

Local conversion and validation files

The dataset-specific script is D:\\code\\scripts\\convert_unitreerobotics_G1_Dex1_Clean_Table.py and the config is D:\\code\\config\\unitreerobotics_G1_Dex1_Clean_Table.yaml. Local validation reports are kept under conversion_reports/; they are not part of the upload set.