THULab/so101_red_cube_bowl_sim_v2

SO101 Red Cube Bowl Sim v2

Apache TsFile conversion of qb1t/so101_red_cube_bowl_sim_v2, a LeRobot simulation trajectory dataset for an SO-101 follower robot.

• Modalities: Time-series
• Robot type: so101_follower
• Sampling rate: 30 fps
• Task: pick up red cube and place in white bowl
• License: Apache-2.0

Dataset Scale

Item	Count
Episodes	50
Trajectory frames / TsFile rows	20,989
Tasks	1
Source camera videos	100
Converted TsFile files	1

All train episodes are merged into data/so101_red_cube_bowl_sim_v2_train.tsfile. Episodes remain independently queryable through the original episode_index and task_index TAG columns.

TsFile Schema

Role	Columns	Type / meaning
TIME	Time	INT64 milliseconds, round(timestamp * 1000)
TAG	episode_index, task_index	Original LeRobot episode and task identifiers
FIELD	frame_index, sample_index	Original frame index and renamed global source index
FIELD	observation_state_0 ... observation_state_5	SO-101 state vector, FLOAT
FIELD	action_0 ... action_5	SO-101 action vector, FLOAT

The six-element state and action vectors use this source order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos.

Conversion Notes

• observation.state[6] and action[6] are flattened into scalar single-precision FLOAT fields. Full source prefixes are retained with . replaced by _.
• Time = round(timestamp * 1000) milliseconds and restarts for every episode. The source timestamp column is dropped because it is redundant with Time.
• Source index is retained as sample_index; frame_index is unchanged.
• All 20,989 authoritative trajectory rows are retained, with no duplicate (episode_index, task_index, Time) keys.
• The source repository contains both episode_*.parquet and file-*.parquet layouts for the same 50 episodes. The actual meta/info.json identifies episode_*.parquet as its data path, and the sampled first pair has identical schema, row count, and values. Only the authoritative episode layout is converted, preventing duplicate rows.
• Source metadata is mirrored. meta/info.json is rewritten so its features and data path describe the converted TsFile while preserving source video details under tsfile_conversion.

Camera Videos

The source camera streams observation.images.ego and observation.images.external_D455 are not included in this repository. They are 640 x 480, 30 fps AV1 streams without audio. The original 100 videos remain available from the source dataset's videos/ tree.

TsFile rows stay aligned with source video frames through episode_index and frame_index.

Reading

from tsfile import TsFileReader

reader = TsFileReader("data/so101_red_cube_bowl_sim_v2_train.tsfile")
print(reader.get_all_table_schemas())
reader.close()

Source

• Original dataset: https://huggingface.co/datasets/qb1t/so101_red_cube_bowl_sim_v2
• Publisher: qb1t
• Created using: https://github.com/huggingface/lerobot