THULab/Aadhavan_so101_bio_final

SO101 Biomedical Waste Sorting TsFile

This dataset provides the numeric robot trajectories from Aadhavan/so101_bio_final as one Apache TsFile table. The LeRobot v2.1 SO101 demonstrations cover picking up biomedical waste and placing it in the respective bin.

Source Dataset

• Author, repository owner, and uploader: Aadhavan Muthurengan (Aadhavan)
• License: Apache-2.0
• Split: train
• Scale: 50 episodes, 22,326 frame rows, 1 task, and 50 source Parquet shards
• Sampling frequency: 30 fps
• Robot type: so101
• LeRobot codebase version: v2.1
• Source data layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Paper and citation: the source card does not provide either one
• Task 0: pick up biomedical waster and put it in the respective bin.

Data Layout

• TsFile: data/aadhavan_so101_bio_final.tsfile
• Table: aadhavan_so101_bio_final
• Rows: 22,326
• Devices: 50, identified by the two TAG columns
• Time precision: milliseconds
• Per-episode Time range: 0 to 29,333 ms

All 50 episode shards are represented in the table, including the six very short source episodes with one or two rows. Filter by episode_index and task_index to select an episode trajectory.

Schema

Column	TsFile type	Role	Meaning
Time	TIMESTAMP	TIME	round(timestamp * 1000) in milliseconds
episode_index	STRING	TAG	Source episode index stored by the TsFile device/tag mechanism
task_index	STRING	TAG	Source task index stored by the TsFile device/tag mechanism
frame_index	INT64	FIELD	Frame position within the episode
sample_index	INT64	FIELD	Source global index value
action_0 ... action_5	FLOAT	FIELD	Six SO101 action components
observation_state_0 ... observation_state_5	FLOAT	FIELD	Six SO101 joint-state components

The action and state component order is main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

Transform Details

• Time is derived from the source timestamp and restarts at zero for every episode. The source timestamp is omitted because it equals Time / 1000 seconds.
• index is renamed to sample_index; frame_index is preserved.
• action[6] is flattened to action_0 through action_5.
• observation.state[6] is flattened to observation_state_0 through observation_state_5.
• Rows are ordered by episode_index, task_index, and Time.
• No source trajectory row, episode, task, action dimension, or state dimension is removed.

Videos

The 100 source AV1 MP4 files remain in the original repository and are not included here. Each camera has 50 files at 640x480 and 30 fps:

• observation.images.front
• observation.images.robocam

Their source pattern is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index and frame_index to align each numeric row with both video streams.

Usage

from tsfile import ColumnCategory, TsFileReader

path = "data/aadhavan_so101_bio_final.tsfile"
reader = TsFileReader(path)
table_name = "aadhavan_so101_bio_final"
schema = reader.get_all_table_schemas()[table_name]
columns = [
    column.get_column_name()
    for column in schema.get_columns()
    if column.get_category() in (ColumnCategory.TAG, ColumnCategory.FIELD)
]

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

Attribution

The demonstrations were published by Aadhavan Muthurengan under the Apache-2.0 license and were created with LeRobot. Cite the original Hugging Face dataset URL above when using the data.