THULab/bjb7_so101_pen_touch_test_1

SO101 Pen Touch Test 1 TsFile

This is an Apache TsFile conversion of bjb7/so101_pen_touch_test_1 (https://huggingface.co/datasets/bjb7/so101_pen_touch_test_1), a LeRobot v2.1 SO101 dataset for Touch the pen. Numeric trajectory data are in one table-model TsFile; videos remain in the original repository.

Source and provenance

• Original author/uploader: Brian Blankenau (bjb7), https://huggingface.co/bjb7
• License: Apache-2.0; no paper or completed citation is supplied by the source card.
• Robot: so101; LeRobot codebase: v2.1; split: train; sampling rate: 30 fps.
• 100 episodes, 21,648 frame rows, one task (task_index=0: Touch the pen.), 100 source Parquet shards, and 200 source MP4 files.
• Source frame path: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet

Converted files

• TsFile: data/so101_pen_touch_test_1.tsfile
• Table: so101_pen_touch_test_1
• Rows: 21,648; devices/episodes: 100
• Original source Parquet bytes (100 shards): 1,105,996; TsFile bytes: 337,278 (30.50% of source, 69.50% smaller).
• The normalized single-file staging Parquet is 327,182 bytes; the small 10,096-byte TsFile difference is table/TAG metadata overhead, not duplicated measurements.

Schema

Time is INT64 milliseconds computed as round(timestamp * 1000) and restarts from zero within each episode. The source timestamp is dropped because it equals Time / 1000 seconds.

TAG/device columns:

• 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

Vector flattening preserves source names: action[6] becomes action_0 ... action_5; observation.state[6] becomes observation_state_0 ... observation_state_5 (FLOAT32). index is renamed to sample_index; frame_index is retained.

Encoding and compression

The generated schema was verified after writing. The configured compact profile is:

• time_encoding: TS_2DIFF
• time_compression: LZ4
• float_encoding: GORILLA
• float_compression: LZ4
• double_encoding: GORILLA
• double_compression: LZ4
• int32_encoding: TS_2DIFF
• int32_compression: LZ4
• int64_encoding: TS_2DIFF
• int64_compression: LZ4
• boolean_encoding: RLE
• boolean_compression: LZ4
• tag_storage: TsFile table-model device/tag columns

TAG values use TsFile table-model device/tag storage. No numeric trajectory rows or state/action dimensions are dropped.

Videos and alignment

Videos are not included in this TsFile repository. The original Hugging Face dataset keeps 200 MP4 files (100 per stream, about 562,260,333 bytes / 536.21 MiB):

• observation.images.camera_2: https://huggingface.co/datasets/bjb7/so101_pen_touch_test_1/tree/main/videos/chunk-000/observation.images.camera_2
• observation.images.camera_4: https://huggingface.co/datasets/bjb7/so101_pen_touch_test_1/tree/main/videos/chunk-000/observation.images.camera_4 Each episode numeric rows remain frame-aligned through episode_index and frame_index.

Validation

Local validation status: PASS. Staged Parquet and TsFile metadata both contain 21,648 rows; the schema exposes 14 FIELD columns and 2 TAG columns. The detailed JSON/Markdown report remains in the local workdir and is not part of the upload set.

Minimal read example

from tsfile import TsFileReader
reader = TsFileReader("data/so101_pen_touch_test_1.tsfile")
with reader.query_table("so101_pen_touch_test_1", ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"], batch_size=1024) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()