Datasets:

azithromycin
/

ac-pate_test_mobile_bimanual_converted

license: apache-2.0
authors:
  - Achal Patel
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
  - robotics
  - lerobot
  - bimanual
  - handover
  - mobile-robot
modality:
  - timeseries
  - tabular
pretty_name: Test Mobile Bimanual Converted (TsFile)
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/ac_pate_test_mobile_bimanual_converted.tsfile
size_categories:
  - 10K<n<100K

Test Mobile Bimanual Converted (TsFile)

This dataset is an Apache TsFile conversion of ac-pate/test_mobile_bimanual_converted, a LeRobot mobile bimanual manipulation dataset. It contains numeric robot actions, states, frame timing, and episode/task tags. Videos remain in the original Hugging Face dataset and are not duplicated.

Source Dataset and Author

Original dataset: ac-pate/test_mobile_bimanual_converted
Original author, repository owner, and uploader: Achal Patel (ac-pate)
License: Apache-2.0
Robot type: mimic_follower
LeRobot codebase version: v3.0
Task (task_index=0): Grasp the blue block from one side using one arm and transfer it to the other arm on the opposite side
Split: train, episodes 0 through 24
Sampling rate: 30 fps
Scale: 25 episodes, 15,177 frames, 1 task
Source Parquet layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
Source Parquet shards: 25
Paper, homepage, and completed citation: not provided by the source card

The pinned Parquet files, v3 feature metadata, and remote file tree are authoritative for this conversion. The source README embeds an older v2.1 description with 12-dimensional vectors and three video streams. The current repository contains 15-dimensional vectors and four streams; the remote tree also corrects the stale total_videos: 75 count and video path order in source meta/info.json to 100 files under videos/{video_key}/chunk-000/.

TsFile Schema

Time is an INT64 millisecond timestamp computed as round(timestamp * 1000). It starts at zero and is strictly increasing within each episode.

Role	Columns	TsFile type
TIME	`Time`	INT64/TIMESTAMP
TAG	`episode_index`, `task_index`	STRING device tags
FIELD	`frame_index`, `sample_index`	INT64
FIELD	`observation_state_0` ... `observation_state_14`	FLOAT
FIELD	`action_0` ... `action_14`	FLOAT

The 15 state/action dimensions retain the source order: six left-arm values, six right-arm values, then three mobile-base values. In the action vector the base values are base_vx, base_vy, and base_omega; in state they are base_x, base_y, and base_theta.

Videos

Videos are not included in this converted repository. The pinned source has 100 frame-aligned AV1 MP4 files at 30 fps, split across four streams:

observation.images.front - 25 per-episode MP4 files
observation.images.left_wrist - 25 per-episode MP4 files
observation.images.right_wrist - 25 per-episode MP4 files
observation.images.top - 25 per-episode MP4 files

Source template: videos/{video_key}/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.mp4. Each stream has one file per episode. Numeric rows remain aligned with the original videos through episode_index, frame_index, and meta/episodes/.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/ac_pate_test_mobile_bimanual_converted.tsfile")
table_name = "ac_pate_test_mobile_bimanual_converted"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "observation_state_0",
    "action_0",
]

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

Citation

The source dataset does not provide a paper or completed citation. Cite Achal Patel, the original Hugging Face dataset, and the pinned source revision when using this converted artifact.