THULab/kyomangold_so100_ttt_B1_v2

SO100 Tic-Tac-Toe B1 v2 TsFile

Apache TsFile edition of kyomangold/so100_ttt_B1_v2, a LeRobot v2.1 SO100 robot dataset for picking up the round orange token and placing it on tic-tac-toe square B1. Numeric trajectories are stored in one table-model TsFile.

Source And Attribution

• Original author and repository owner: Kyo Mangold (kyomangold)
• License: Apache-2.0
• Task: Pick up the round, orange token and place it in the grid on square B1.
• Split: train; 30 episodes, 15,240 frames, 1 task, sampled at 30 FPS, from 30 source Parquet shards.
• The source card does not provide a homepage, paper, or completed BibTeX citation.

TsFile Data

The file data/kyomangold_so100_ttt_b1_v2.tsfile contains 15,240 rows across 30 episode/task devices. Time is round(timestamp * 1000) in milliseconds and restarts at zero for each episode. The source timestamp field is omitted because it is equivalent to Time / 1000 seconds.

Column	Role	Type	Source mapping
Time	TIME	INT64 milliseconds	round(timestamp * 1000)
episode_index, task_index	TAG	STRING	Original INT64 identifiers
frame_index	FIELD	INT64	Preserved frame position
sample_index	FIELD	INT64	Renamed from 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]

Vector prefixes are preserved and dots are replaced with underscores. The six dimensions retain the source order: main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper. No trajectory rows, episodes, tasks, action dimensions, or state dimensions are removed.

Source Videos

Videos are not included in this repository. They remain in the original Hugging Face dataset under videos/chunk-000. The two source streams are observation.images.laptop and observation.images.phone, with 30 MP4 files per stream. The source template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4; use episode_index and frame_index to align each numeric row with the corresponding 30 FPS frames.

Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/kyomangold_so100_ttt_b1_v2.tsfile")
with reader.query_table(
    "kyomangold_so100_ttt_b1_v2",
    ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"],
    batch_size=1024,
) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()