THULab/Bartm3_dice4

Dice4 TsFile

This Apache TsFile dataset is based on Bartm3/dice4, a LeRobot v2.1 SO100 robot dataset containing demonstrations for putting a dice or tape in a bin.

Source

• Repository owner and publisher: Mathias (Bartm3)
• License: Apache-2.0
• Paper or formal citation: not supplied by the source dataset card
• Robot type: so100
• Split: train
• Sampling rate: 30 fps
• Scale: 108 episodes, 55,590 frame rows, 2 tasks, 108 source data files
• Source data layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Task index 0 is "Grasp a dice and put it in the bin." (90 episodes, 46,502 rows). Task index 1 is "Grasp a tape and put it in the bin." (18 episodes, 9,088 rows).

The current meta/info.json and all source data files provide these counts. The smaller 10-episode JSON example embedded in the source card is stale.

TsFile Data

• Path: data/bartm3_dice4.tsfile
• Table: bartm3_dice4
• Rows: 55,590
• Episode/task devices: 108
• Time precision: milliseconds
• Source numeric-data size: 3,356,504 bytes
• TsFile size: 1,214,202 bytes

The train split is one table. Use episode_index and task_index as TAG filters to select an episode or task.

Schema

Time = round(timestamp * 1000) milliseconds and restarts at zero in every episode. Source timestamp is omitted because it equals Time / 1000 seconds.

Column	TsFile role	Type	Source mapping
Time	TIME	INT64	round(timestamp * 1000)
episode_index	TAG	STRING device segment	Source episode_index
task_index	TAG	STRING device segment	Source task_index
frame_index	FIELD	INT64	Preserved
sample_index	FIELD	INT64	Source index
action_0 ... action_5	FIELD	FLOAT	Flattened from action[6]
observation_state_0 ... observation_state_5	FIELD	FLOAT	Flattened from observation.state[6]

The six action and state dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper. All source rows, episodes, tasks, action dimensions, and state dimensions are retained.

Videos

Videos are not included here. The original repository stores two frame-aligned 640x480 AV1 MP4 streams at 30 fps without audio:

• observation.images.laptop: 108 files, 246,699,809 bytes
• observation.images.phone: 108 files, 428,651,369 bytes

Together they contain 216 files totaling 675,351,178 bytes. episode_index and frame_index preserve alignment with each original per-episode file.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/bartm3_dice4.tsfile")
with reader.query_table(
    "bartm3_dice4",
    ["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()

Citation

The source card does not provide a paper or completed citation. Cite the original Hugging Face repository and its owner/publisher, Mathias (Bartm3).