azithromycin/unitreerobotics_G1_WBT_Dex1_Put_Clothes_into_Washing_Machine

G1 WBT Dex1 Put Clothes into Washing Machine TsFile

This repository is an Apache TsFile conversion of unitreerobotics/G1_WBT_Dex1_Put_Clothes_into_Washing_Machine, a LeRobot v3 robot-manipulation dataset for a Unitree G1 with Dex1 hands.

The converted repository contains numeric robot states, actions, frame timing, episode/task tags, and source metadata. Camera videos remain in the original Hugging Face dataset and are not duplicated here.

Source dataset and attribution

• Source: Unitree G1 WBT Dex1 Put Clothes into Washing Machine
• Source organization: Unitree Robotics (unitreerobotics)
• Uploader shown in the repository and video history: karthus198
• License: Apache-2.0
• Framework: LeRobot v3.0
• Robot type: unitree_g1
• Task: put clothes into a washing machine using a Unitree G1 with Dex1 hands
• Paper/citation: the source card does not provide a paper or formal citation

Source scale and split

Item	Value
Source split	train
Source Parquet shards	1
Source data layout	data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
Observed Parquet rows	124,184
Observed episodes	155 (episode indexes 0–154)
Tasks	1 (task index 0)
Sampling rate	30 fps
Converted TsFiles	1

The source meta/info.json declares 154 episodes, 123,603 frames, and train: 0:154, while the Parquet shard contains an additional episode 154 with 581 rows. The conversion preserves all 124,184 Parquet rows and all 155 observed episodes. The original declaration is retained in metadata as a documented source inconsistency.

Converted file

• TsFile: data/unitreerobotics_g1_wbt_dex1_put_clothes_into_washing_machine.tsfile
• Table: unitreerobotics_g1_wbt_dex1_put_clothes_into_washing_machine
• Rows: 124,184
• Devices/TAG combinations: 155
• Time precision: milliseconds
• File size: 47,175,737 bytes
• Source Parquet size: 71,575,199 bytes
• Size ratio: 65.91% of the source Parquet

TsFile schema

The logical Time value is round(timestamp * 1000) milliseconds. It restarts for every episode. The Python reader exposes the physical time column as lowercase time; it is the TsFile TIME column, not a normal FIELD.

Source / converted column	TsFile type	Role	Notes
timestamp → Time	TIMESTAMP / INT64	TIME	Milliseconds; source timestamp is dropped after conversion
episode_index	STRING	TAG	Source INT64 stored as a TsFile device/tag segment
task_index	STRING	TAG	Source INT64 stored as a TsFile device/tag segment
frame_index	INT64	FIELD	Preserved
index → sample_index	INT64	FIELD	Renamed to avoid ambiguity
observation.state.ee_state[12]	FLOAT	FIELD	observation_state_ee_state_0 … _11
observation.state.hand_state[2]	FLOAT	FIELD	observation_state_hand_state_0 … _1
observation.state.robot_q_current[36]	FLOAT	FIELD	observation_state_robot_q_current_0 … _35
action.ee_action[12]	FLOAT	FIELD	action_ee_action_0 … _11
action.hand_cmd[2]	FLOAT	FIELD	action_hand_cmd_0 … _1
action.robot_q_desired[36]	FLOAT	FIELD	action_robot_q_desired_0 … _35

Vector columns are flattened in row-major order. The full source name is preserved by replacing each dot with an underscore and appending the element index. The resulting table has 102 FIELD columns, 2 TAG columns, and 1 TIME column.

Encoding and compression

Data type	Encoding	Compression
FLOAT, DOUBLE	GORILLA	LZ4
INT32, INT64	TS_2DIFF	LZ4
TIME	TS_2DIFF	LZ4
BOOLEAN	RLE	LZ4
TAG	TsFile table device/tag mechanism	TsFile-supported TAG storage

This dataset contains FLOAT and INT64 FIELD data and no BOOLEAN FIELD columns. The writer is configured explicitly before table creation; TAG values are stored as STRING device segments while their source INT64 types are recorded in meta/info.json.

Dropped and omitted content

• timestamp is dropped only after generating Time, because it is redundant with Time / 1000 seconds.
• No numeric rows, state/action dimensions, frame_index, episode indexes, or task indexes are dropped.
• The four video features are omitted from the TsFile and remain in the source repository.

Videos

The original video tree is approximately 8.41 GB and follows videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. It contains 43 physical MP4 shards:

• observation.images.head_stereo_left: 10 files, file-000.mp4–file-009.mp4
• observation.images.head_stereo_right: 13 files, file-000.mp4–file-012.mp4
• observation.images.wrist_left: 10 files, file-000.mp4–file-009.mp4
• observation.images.wrist_right: 10 files, file-000.mp4–file-009.mp4

Numeric rows remain aligned with the original videos through episode_index, frame_index, and the per-episode offsets in meta/episodes/.

Validation

Local validation confirmed:

• TsFile metadata rows: 124,184
• Query readback rows: 124,184
• TAG/device combinations: 155
• Time is monotonic within every episode after TAG/Time sorting
• 102 FIELD + 2 TAG + 1 TIME columns are present
• The generated TsFile is smaller than the source Parquet

The detailed local files validation_report.json and VALIDATION.md are conversion artifacts and should not be uploaded with the dataset.

Usage

from tsfile import TsFileReader

path = "data/unitreerobotics_g1_wbt_dex1_put_clothes_into_washing_machine.tsfile"
table_name = "unitreerobotics_g1_wbt_dex1_put_clothes_into_washing_machine"

reader = TsFileReader(path)
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_ee_action_0",
    "observation_state_ee_state_0",
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())