THULab/berkeley_autolab_ur5

Berkeley Autolab UR5 TsFile

This dataset is an Apache TsFile conversion of lerobot/berkeley_autolab_ur5, the Berkeley UR5 Demonstration Dataset published in LeRobot format.

Modalities: Time-series. The converted repository contains numeric robot observations, actions, frame timing, reward/done flags, task/episode tags, and source metadata. Camera videos remain in the original Hugging Face dataset.

Source Dataset

• Source dataset: lerobot/berkeley_autolab_ur5
• Homepage: Berkeley UR5 Demonstration Dataset
• Citation authors: Lawrence Yunliang Chen, Simeon Adebola, Ken Goldberg
• License: CC-BY-4.0
• Robot type: unknown
• LeRobot codebase version in downloaded meta/info.json: v3.0
• Split: train
• Scale: 1,000 episodes, 97,939 frames, 5 tasks, 5 fps
• Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
• Source video layout: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4

Task metadata in the downloaded episode file lists these task descriptions:

• 0: sweep the green cloth to the left side of the table
• 1: put the ranch bottle into the pot
• 2: pick up the blue cup and put it into the brown cup.
• 3: take the tiger out of the red bowl and put it in the grey bowl
• 4: put the marker into the bowl

Converted Files

• TsFile: data/berkeley_autolab_ur5.tsfile
• Table: berkeley_autolab_ur5
• Rows: 97,939
• Episodes: 1,000
• Tasks: 5
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json updated to describe the converted TsFile artifact.

Schema

Time is synthesized as round(timestamp * 1000) milliseconds and restarts within each episode.

TAG columns:

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index, renamed from the source index
• next_reward, renamed from the source next.reward
• next_done, renamed from the source next.done

Flattened FLOAT FIELD groups:

• observation.state[8] -> observation_state_0 ... observation_state_7
• action[7] -> action_0 ... action_6

Conversion Notes

• The generic lerobot converter was used.
• The numeric frame Parquet shard is merged into one TsFile table. Use episode_index and task_index TAG filters to select an episode or task.
• Vector columns are flattened to scalar TsFile fields. Source column prefixes are preserved, with . replaced by _.
• The source timestamp column is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• The source index column is renamed to sample_index; next.reward and next.done are renamed to next_reward and next_done.
• Source camera video streams are not included in this repository: observation.images.hand_image, observation.images.image, and observation.images.image_with_depth. They remain available in the original dataset under videos/.
• The current frame Parquet schema downloaded for this conversion does not include a language_instruction column. The converted schema follows the available Parquet and meta/info.json files from the source repository.
• No rows and no numeric time-series fields are intentionally dropped other than the redundant timestamp column noted above.

Validation

The converted TsFile was read back with the Apache TsFile Python SDK. The TsFile metadata row count and query readback both matched the staged Parquet: 97,939 rows.

Usage

from tsfile import TsFileReader

path = "data/berkeley_autolab_ur5.tsfile"
reader = TsFileReader(path)

schemas = reader.get_all_table_schemas()
table_name = "berkeley_autolab_ur5"
columns = [
    column.get_column_name()
    for column in schemas[table_name].get_columns()
    if column.get_column_name() not in {"Time"}
]

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

Citation

@misc{BerkeleyUR5Website,
  title = {Berkeley {UR5} Demonstration Dataset},
  author = {Lawrence Yunliang Chen and Simeon Adebola and Ken Goldberg},
  howpublished = {https://sites.google.com/view/berkeley-ur5/home},
}