README.md

metadata

license: apache-2.0
authors:
  - Adrien Payan
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
  - robotics
  - lerobot
  - so100
  - manipulation
modality:
  - timeseries
  - tabular
pretty_name: SO100 Ball8 TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/apayan_so100_ball8.tsfile
size_categories:
  - 100K<n<1M

SO100 Ball8 TsFile

Apache TsFile conversion of apayan/so100_ball8, a LeRobot v2.1 SO100 dataset for "Grasp object and put it in the box." Numeric observations, actions, timing, and episode/task metadata are included; camera videos remain upstream.

Source facts

• Author/uploader: Adrien Payan
• License: Apache-2.0; pinned revision: f318cc9490b327bd1382e29c528f7d7348aa4716
• Robot: so100; split: train; sampling rate: 30 fps; task index 0
• 104 episodes, 110,416 frames, 1 task, 104 source Parquet shards
• Source layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Current metadata is authoritative; source README embeds a stale 5-episode/5,992-frame example.

Converted artifact

• TsFile: data/apayan_so100_ball8.tsfile; table: apayan_so100_ball8
• Rows: 110,416; TAG devices: 104; size: 1.14 MiB (source Parquet: 4.58 MiB)
• meta/ is mirrored; meta/info.json documents the TsFile mapping and video policy.

Schema and conversion

Time = round(timestamp * 1000) integer milliseconds, restarting per episode. Source timestamp is dropped because it equals Time / 1000 seconds. Rows are sorted by episode_index, task_index, Time; frame_index is retained.

TAG/device columns: episode_index, task_index (source INT64, native TsFile TAG/device mechanism).

FIELD columns: frame_index, sample_index (renamed from index), action_0..action_5, and observation_state_0..observation_state_5 (FLOAT32). No numeric rows, episodes, tasks, or vector dimensions are dropped.

Encoding/compression: FLOAT/DOUBLE -> GORILLA + LZ4; INT32/INT64 -> TS_2DIFF + LZ4; Time -> TS_2DIFF + LZ4; BOOLEAN -> RLE + LZ4 (no BOOLEAN source fields).

Videos

Videos are not included here. The source has 312 MP4 files (104 episodes x three 30-fps streams):

• observation.images.front
• observation.images.top
• observation.images.wrist

Template: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index and frame_index to align rows with source video frames; meta/episodes.jsonl is mirrored.

Validation

All 104 Parquet schemas matched and the staged table exactly matched the independent transformation. TsFile inspection found one table, 2 TAG columns, 14 FIELD columns, and 110,416 metadata rows; query readback returned 110,416 rows. Reports remain local at D:/code/workdir/apayan_so100_ball8/validation_report.json and D:/code/workdir/apayan_so100_ball8/VALIDATION.md.

Minimal read example

from tsfile import TsFileReader
reader = TsFileReader("data/apayan_so100_ball8.tsfile")
columns = ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"]
with reader.query_table("apayan_so100_ball8", columns, batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()

The source card provides no paper or completed citation; cite the original Hugging Face dataset and Adrien Payan (apayan).