THULab/bjb7_so101_pen_touch_test_2

SO101 Pen Touch Test 2 TsFile

This dataset is an Apache TsFile conversion of bjb7/so101_pen_touch_test_2, a LeRobot v2.1 SO101 robot-manipulation dataset containing demonstrations for the task "Touch the pen."

Modalities: Time-series, Tabular. The converted repository contains numeric robot states, actions, frame timing, and episode/task tags. The two camera streams remain in the original Hugging Face dataset and are linked below.

Source Dataset and Provenance

• Original dataset: bjb7/so101_pen_touch_test_2
• Pinned source revision: 7e8dd963ae6e496b5b6e8d2200cac4a14ed5a9c5
• Original repository creator and uploader: Brian Blankenau (bjb7)
• License: Apache-2.0
• Robot type: so101
• LeRobot codebase version: v2.1
• Task: Touch the pen. (task_index = 0)
• Split: train
• Sampling rate: 30 fps
• Scale: 100 episodes, 24,167 frame rows, 1 task, 100 source Parquet files, 200 source videos
• Source frame layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

The source dataset card provides no paper or completed citation. The Hugging Face repository history attributes all source commits to bjb7; the associated profile identifies the account as Brian Blankenau.

Converted Files

• TsFile: data/so101_pen_touch_test_2_train.tsfile
• Table: so101_pen_touch_test_2_train
• Rows: 24,167
• Episodes/devices: 100
• TsFile size: 388,152 bytes (0.37 MiB)
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten to describe the TsFile artifact and conversion mapping.

TsFile Schema

Time is an INT64 millisecond timestamp computed as round(timestamp * 1000) and restarts for each episode.

TAG columns (stored as TsFile STRING tags while preserving the original source dtype in conversion metadata):

• episode_index
• task_index

FIELD columns:

• frame_index
• sample_index
• action_0
• action_1
• action_2
• action_3
• action_4
• action_5
• observation_state_0
• observation_state_1
• observation_state_2
• observation_state_3
• observation_state_4
• observation_state_5

Flattened vector groups:

• action -> action_0 ... action_5 (6 FLOAT fields)
• observation.state -> observation_state_0 ... observation_state_5 (6 FLOAT fields)

Conversion Notes

• The shared config-driven lerobot converter is used; the included convert_so101_pen_touch_test_2.py is the dataset-specific local entry point.
• The train split is merged into one table-model TsFile. Filter by episode_index and task_index to select an episode or task.
• Storage profile: Time uses TS_2DIFF + LZ4; FLOAT/DOUBLE use GORILLA + ZSTD; INT32/INT64 use TS_2DIFF + ZSTD; BOOLEAN uses RLE + ZSTD. The episode_index and task_index columns remain TsFile table-model TAG/device columns.
• action[6] and observation.state[6] are flattened to scalar FLOAT fields; the full source prefix is retained and . is replaced with _.
• The source timestamp column is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• The source index column is retained as sample_index; frame_index is retained unchanged.
• All 24,167 source rows and all 12 action/state dimensions are retained.

Videos

Videos are not duplicated in this converted repository. The pinned source contains two frame-aligned camera streams, each with 100 per-episode MP4 files:

• observation.images.camera_2 - 100 per-episode MP4 files
• observation.images.camera_4 - 100 per-episode MP4 files

Together the 200 MP4 files occupy 620,609,875 bytes (about 591.86 MiB). The numeric TsFile rows remain aligned with the original videos through episode_index, frame_index, and the source per-episode metadata.

Validation

The generated TsFile is checked for successful tool completion, non-zero size, table schema, metadata row-count equality with the staged Parquet, and a query readback sample. See VALIDATION.md and validation_report.json.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/so101_pen_touch_test_2_train.tsfile")
table_name = "so101_pen_touch_test_2_train"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_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 card provides no paper or completed citation. Cite the original Hugging Face dataset and Brian Blankenau (bjb7) when using this converted artifact.