THULab/jccj_so100_block_in_cup

SO100 Block in Cup TsFile

This dataset is an Apache TsFile conversion of jccj/so100_block_in_cup, a LeRobot v2.1 SO100 robot-manipulation dataset. It contains numeric trajectories, timing, episode/task tags, and source metadata. Videos remain in the original Hugging Face repository.

Source Dataset and Attribution

• Original dataset: jccj/so100_block_in_cup
• Pinned revision: 87a2b14997c2fabe117ca2b1fa4e7a8a42356623
• Original author/uploader: JCCJ (jccj)
• Authorship evidence: the repository lists one contributor (jccj) across three commits; the source card does not provide a separate author list, paper, or completed citation.
• License: Apache-2.0
• Task: “Grasp a block and put it in the cup.”
• Robot: so100; LeRobot version: v2.1
• Split: train; sampling rate: 30 fps
• Scale: 50 episodes, 21,782 frames, 1 task
• Source shards: 50 Parquet files under data/chunk-000/episode_{episode_index:06d}.parquet

Converted File

• TsFile: data/jccj_so100_block_in_cup.tsfile (318,964 bytes)
• Table: jccj_so100_block_in_cup
• Rows: 21,782; episodes/devices: 50; tasks: 1
• Time precision: milliseconds
• TsFile/source-Parquet size ratio: 0.298
• meta/ is preserved, with meta/info.json updated for the TsFile artifact.

Schema

Time = round(timestamp * 1000) milliseconds. Time restarts at zero in every episode and source timestamp is then dropped because it equals Time / 1000.

TAG columns (TsFile device/tag mechanism):

• 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. Dots in source vector names are replaced by underscores. No numeric row, episode, task, or vector dimension is dropped.

Encodings and Compression

• FLOAT/DOUBLE: GORILLA + LZ4
• INT32/INT64: TS_2DIFF + LZ4
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4 (the source has no BOOLEAN field)
• TAG: TsFile table/device TAG storage

The on-disk table schema, Time codec, every physical FIELD codec, and all 21,782 rows were read back with the Apache TsFile Java API.

Videos

Videos are not included in this TsFile repository. The original dataset has 100 frame-aligned AV1 MP4 files (415,886,178 bytes, about 396.5 MiB), 640x480, 30 fps, no audio, in two streams:

• observation.images.phone
• observation.images.wrist

Each stream uses videos/chunk-000/{video_key}/episode_{episode_index:06d}.mp4. episode_index, frame_index, and meta/episodes.jsonl preserve alignment.

Validation

Source, staged Parquet, and complete Java TsFile readback all contain 21,782 rows. TAG values, 50 episode indexes, one task, vector widths, zero-based monotonic Time per episode, physical codecs, non-zero size, and SHA-256 were checked locally. Conversion scripts and validation reports are intentionally not included in this upload-ready directory.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/jccj_so100_block_in_cup.tsfile")
with reader.query_table(
    "jccj_so100_block_in_cup",
    ["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())
reader.close()