Datasets:

jzpeng
/

Xense_whiteboard_inspect_0708

license: apache-2.0
authors:
  - XenseRobotics
  - fza
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
modality:
  - timeseries
  - tabular
pretty_name: Whiteboard Inspect 0708 TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/xense_whiteboard_inspect_0708.tsfile
size_categories:
  - 10K<n<100K

Whiteboard Inspect 0708 TsFile

Apache TsFile edition of Xense/whiteboard_inspect_0708, a LeRobot v3.0 bimanual Flexiv Rizon 4 robotics dataset.

Source and attribution

Original dataset: https://huggingface.co/datasets/Xense/whiteboard_inspect_0708
Publishing organization: XenseRobotics (Xense)
Repository contributor: fza (fza1796262052)
License: Apache-2.0
Task: Pick up the inspection probe, press it on each red-marked spot on the whiteboard one by one, put down the probe, pick up the eraser, and wipe off all the red marks.
Robot: bi_flexiv_rizon4_rt
Train split: 19 episodes, 67,684 frame rows, one task, 30 fps, 19 Parquet files
Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet

The source dataset card does not provide a paper or BibTeX citation. Cite the original Hugging Face dataset, XenseRobotics, and contributor fza when using this data.

TsFile schema

Time = round(timestamp * 1000) as INT64 milliseconds and restarts at zero in each episode. The source timestamp column is removed because it is exactly represented by Time / 1000 seconds. index is renamed to sample_index, and frame_index is retained.

Columns	TsFile type	Role
`Time`	INT64/TIMESTAMP	TIME
`episode_index`, `task_index`	STRING device segments	TAG
`frame_index`, `sample_index`	INT64	FIELD
`action_0` ... `action_19`	FLOAT	FIELD
`observation_state_0` ... `observation_state_19`	FLOAT	FIELD

The 20 action and 20 state elements follow the source feature order. Each arm contains TCP x/y/z, six rotation representation values, r1 through r6, in the source-defined order, and a gripper position value. The left-arm values precede the right-arm values. Dots in source vector names are represented by the scalar field prefix and element index. No numeric rows, action dimensions, or state dimensions are omitted.

Storage uses GORILLA + LZ4 for FLOAT/DOUBLE, TS_2DIFF + LZ4 for INT32/INT64, TS_2DIFF + LZ4 for Time, and RLE + LZ4 for BOOLEAN fields. TAG values use the TsFile table/device mechanism.

Original videos

The 133 source MP4 files remain at videos/ and are not included here. The path template is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 for these streams:

observation.images.head (640 x 480, H.264, 30 fps)
observation.images.left_wrist (640 x 480, H.264, 30 fps)
observation.images.right_wrist (640 x 480, H.264, 30 fps)
observation.images.left_tactile_0 (700 x 400, H.264, 30 fps)
observation.images.left_tactile_1 (700 x 400, H.264, 30 fps)
observation.images.right_tactile_0 (700 x 400, H.264, 30 fps)
observation.images.right_tactile_1 (700 x 400, H.264, 30 fps)

Use episode_index and frame_index to align numeric rows with the matching frame in each original per-episode video.

Read example

from tsfile import TsFileReader

reader = TsFileReader("data/xense_whiteboard_inspect_0708.tsfile")
with reader.query_table(
    "xense_whiteboard_inspect_0708",
    ["episode_index", "task_index", "Time", "frame_index", "action_0"],
    batch_size=4096,
) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()