THULab/slowturtle99_so100_pick_place

SO100 Pick Place TsFile

This dataset is an Apache TsFile conversion of slowturtle99/so100_pick_place (https://huggingface.co/datasets/slowturtle99/so100_pick_place), a LeRobot v2.1 SO100 robot-manipulation dataset.

Modalities: Time-series. It contains numeric observations, actions, frame timing, episode/task tags, and mirrored source metadata. Camera videos remain in the original Hugging Face dataset.

Source Dataset and Author

• Source dataset: slowturtle99/so100_pick_place (https://huggingface.co/datasets/slowturtle99/so100_pick_place)
• Pinned revision: e82ce5fd2bd8448c005dddebcf819123b1ff7ac2 (https://huggingface.co/datasets/slowturtle99/so100_pick_place/tree/e82ce5fd2bd8448c005dddebcf819123b1ff7ac2)
• Original author/uploader: slowturtle99 (https://huggingface.co/slowturtle99)
• Authorship evidence: all five source commits are attributed to slowturtle99; no separate formal author list or paper is supplied.
• License: Apache-2.0; robot type: so100; LeRobot version: v2.1; split: train; sampling rate: 30 fps
• Scale: 120 episodes, 37,162 frames, 1 task
• Episode length: 226-407 frames (median 307.5)
• Source frame layout: data/chunk-000/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-000/observation.images.webcam/episode_{episode_index:06d}.mp4

The source README embeds an outdated 40-episode/11,629-frame info.json excerpt. The pinned repository files used here contain 120 episodes and 37,162 frames.

Converted Files

• TsFile: data/slowturtle99_so100_pick_place_train.tsfile (852,976 bytes)
• Table: slowturtle99_so100_pick_place_train
• Rows: 37,162; episodes/devices: 120; tasks: 1
• Time precision: milliseconds
• meta/ is mirrored from the source, with meta/info.json rewritten for the TsFile artifact and source-video policy.

TsFile Schema

Time = round(timestamp * 1000) milliseconds, restarting at zero per episode.

TAG columns:

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index (renamed from source index)

Flattened FLOAT FIELD groups:

• action[6] -> action_0 ... action_5
• observation.state[6] -> observation_state_0 ... observation_state_5

The six dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

Encoding and Conversion Notes

• The shared config-driven lerobot converter is used; this local script is the reproducible dataset-specific entry point.
• All train rows are merged into one table-model TsFile. TAGs use the TsFile device/tag mechanism.
• FLOAT/DOUBLE use GORILLA + LZ4. INT32/INT64 and Time use TS_2DIFF + LZ4. No BOOLEAN fields occur in this source; the requested BOOLEAN profile is RLE + LZ4.
• Vectors are flattened to scalar fields with dots replaced by underscores.
• timestamp is dropped after Time synthesis because it equals Time / 1000; index becomes sample_index; frame_index is kept. No other row or state/action dimension is dropped.

Videos

Videos are not copied here. The pinned source contains 120 frame-aligned MP4 files (139,843,079 bytes, about 133.4 MiB) under https://huggingface.co/datasets/slowturtle99/so100_pick_place/tree/e82ce5fd2bd8448c005dddebcf819123b1ff7ac2/videos/chunk-000/observation.images.webcam. The source metadata describes 640x480 H.264 at 30 fps without audio. episode_index, frame_index, and meta/episodes preserve alignment.

Validation

Apache TsFile SDK metadata and a complete batched query readback both match the staged Parquet (37,162 rows). Source schemas and vector widths were also checked. Exact hashes and checks are in VALIDATION.md and validation_report.json.

Usage

from tsfile import TsFileReader
reader = TsFileReader("data/slowturtle99_so100_pick_place_train.tsfile")
with reader.query_table("slowturtle99_so100_pick_place_train",
        ["episode_index", "task_index", "frame_index", "sample_index",
         "action_0", "observation_state_0"],
        batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())