THULab/samsitol_so100_PnP

SO100 PnP (TsFile)

Source: samsitol/so100_PnP, pinned revision eb59ca16fedd0f7214928d3aca137c5b46d1f8c7.

This is an Apache TsFile conversion of a LeRobot v2.1 SO100 manipulation dataset for the task: Grasp the yellow block and put it in the green bin.

• Modalities: Time-series
• Split: train
• Sampling rate: 30 fps
• Scale: 50 episodes, 23,377 frames, 1 task, 50 source episode Parquet files
• Converted layout: one TsFile with 23,377 rows
• Source repository owner/uploader and author of all source commits: samsitol
• License: Apache-2.0

Converted files

• TsFile: data/so100_pnp_train.tsfile
• Table: so100_pnp_train
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten for the TsFile artifact.

TsFile schema

Role	Columns	Representation
Time	Time	INT64 milliseconds; round(timestamp * 1000); restarts per episode
TAG	episode_index, task_index	Source episode and task dimensions
FIELD	frame_index, sample_index	Source scalar frame identifiers; sample_index is renamed from index
FIELD	action_0 ... action_5	Six FLOAT action values
FIELD	observation_state_0 ... observation_state_5	Six FLOAT robot-state values

The six action/state elements are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

Conversion notes

• The shared config-driven lerobot converter was used.
• All 50 train episodes are merged into one table-model TsFile. Filter by episode_index and task_index to select an episode or task.
• The source timestamp column is omitted because it is exactly represented by Time / 1000 seconds. frame_index is retained.
• The source index column is renamed to sample_index.
• action[6] and observation.state[6] are flattened to scalar float32 fields; every vector element and every source row is retained.
• No numeric rows or numeric dimensions are intentionally dropped.

Encoding and compression

The TsFile was re-encoded with a type-specific compact policy so the converted file remains substantially smaller than the source representation:

• FLOAT/DOUBLE fields: GORILLA encoding + LZ4 compression
• INT32/INT64 fields: TS_2DIFF encoding + LZ4 compression
• Time: TS_2DIFF encoding + LZ4 compression
• BOOLEAN fields: RLE encoding + LZ4 compression
• TAG columns: TsFile table TAG/device mechanism; episode_index and task_index remain source TAG columns

The encoded TsFile is approximately 418 KiB (427,742 bytes) for all 23,377 rows, compared with the previous 1,615,335-byte output.

Videos

Videos are not duplicated in this converted repository. The original dataset contains 100 frame-aligned MP4 files (about 1.12 GiB) in two streams:

• observation.images.head
• observation.images.wrist

The source path template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Numeric rows align with the original videos through episode_index, frame_index, and meta/episodes.jsonl.

Validation

The conversion checks successful TsFile generation, non-zero file size, source and staged row-count equality (23,377 rows), unique (episode_index, task_index, Time) keys, and the converted metadata mapping. If the Apache TsFile Python SDK is installed, the script also opens the file and checks its table schema and metadata row count.

Reading

from tsfile import TsFileReader

path = "data/so100_pnp_train.tsfile"
reader = TsFileReader(path)
table = "so100_pnp_train"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]
with reader.query_table(table, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())

Source and license

The source card states that the dataset was created with LeRobot and is licensed under Apache-2.0. It does not provide a paper, homepage, formal author list, or completed citation. Repository history attributes all source commits to samsitol.