tutorial-ball-2 (LeRobot) — TsFile

This dataset is a lossless conversion to the Apache TsFile format of the HuggingFace LeRobot dataset notmahi/tutorial-ball-2: a low-dimensional robot tutorial trajectory dataset (no video).

Original dataset

• Source dataset: notmahi/tutorial-ball-2
• Format: early LeRobot format (meta_data/ + safetensors)
• Content: purely numeric low-dimensional state/action trajectories — 314,074 frames / 751 episodes / 30 fps. No images or video (meta_data/info.json: video=0).

What is in this repository

data/
└── tutorial_ball_2.tsfile     # numeric time-series (converted)
meta_data/
├── info.json                  # original fps/video flags + tsfile_conversion notes
├── stats.safetensors          # per-feature statistics (copied verbatim)
└── episode_data_index.safetensors  # episode boundaries (copied verbatim)

TsFile storage mapping (table model)

Role	Column(s)	Type	Notes
TAG	episode_id	STRING	episode_{episode_index}, 751 devices (one per episode)
Time	round(frame_index * 1000 / 30) ms	INT64 (ms)	30 fps; frame_index restarts at 0 each episode
FIELD	state_0 … state_3	FLOAT	observation.state[4] expanded
FIELD	action_0, action_1	FLOAT	action[2] expanded
FIELD	episode_index, frame_index, sample_index	INT64	indices (index → sample_index)
FIELD	episode_timestamp_s	FLOAT	(timestamp)
FIELD	next_done	BOOLEAN	(next.done)

Conversion notes

• Purely numeric — the source has no images or video, so only data/ is converted; nothing else needed.
• TAG = episode_id (751 devices). Time = round(frame_index × 1000/30) ms. Because frame_index restarts at 0 within each episode and is strictly increasing, and round(k × 1000/30) is also strictly increasing in k (step ≥ 33 ms), every device's time axis is strictly increasing — no de-duplication or offset needed. (30 fps gives a ~33.333 ms frame interval; with millisecond precision the per-frame times are 0, 33, 67, 100, … — consecutive and collision-free.)
• Array columns expanded: observation.state[4] → state_0..state_3, action[2] → action_0..action_1 (FLOAT, matching the source float32).
• Column names with dots made TsFile-safe (next.done → next_done, …).
• No columns dropped, no rows dropped: all 314,074 frames preserved.
• meta_data/ (info / stats / episode index) is copied over; info.json gains a tsfile_conversion block describing the table layout.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/tutorial_ball_2.tsfile")
schemas = reader.get_all_table_schemas()
tname = next(iter(schemas))

cols = ["episode_id", "state_0", "state_1", "action_0", "action_1"]
with reader.query_table(tname, cols, batch_size=65536) as rs:
    while (batch := rs.read_arrow_batch()) is not None:
        df = batch.to_pandas()
        # ... process ...
reader.close()

Citation

@misc{tutorial_ball_2,
  title  = {tutorial-ball-2 (LeRobot)},
  author = {notmahi},
  url    = {https://huggingface.co/datasets/notmahi/tutorial-ball-2},
  publisher = {Hugging Face}
}

The source HuggingFace dataset does not declare an explicit license.