THULab/andlyu_bb_randomgrab_1_aug_train

bb_randomgrab_1_aug_train (TsFile)

Apache TsFile version of andlyu/bb_randomgrab_1_aug_train, a LeRobot robot manipulation dataset created with LeRobot.

Modalities: Time-series. The original dataset also contains robot camera videos; those videos are not included in this converted repository and remain available from the original Hugging Face dataset.

Overview

A LeRobot v2.1 SO-100 follower dataset recording random-grasp demonstrations. Each row is one frame of a robot episode, with the robot's joint state and the commanded action recorded as fixed-size vectors.

• Scale: 590 episodes, 136,540 frames, 1 task.
• Sampling rate: 30 fps.
• Robot: so100_follower.

Schema (TsFile structure)

• Time (INT64, milliseconds) — round(timestamp * 1000), restarting for each episode.
• episode_index, task_index (TAG, INT64) — episode and task device dimensions; query one episode with WHERE episode_index=0.
• frame_index, sample_index (FIELD, INT64) — frame position and source row index (source index renamed to sample_index).
• observation_state_0..5 (FIELD, FLOAT) — flattened observation.state (6 joints).
• action_0..5 (FIELD, FLOAT) — flattened action (6 commanded motors).

The source timestamp column is represented by Time and not retained as a separate field. Vector columns are flattened to scalar FLOAT fields with their full source-column prefix and an element index. The camera video streams are not included; they remain available in the original dataset.

Usage

Read data/bb_randomgrab_1_aug_train.tsfile with the Apache TsFile Java or Python SDK. All episodes share one table; use the TAG columns to select an episode or task.

Source & license

• Original dataset: https://huggingface.co/datasets/andlyu/bb_randomgrab_1_aug_train
• Author / publisher: andlyu
• Paper: not declared by the original dataset
• License: not declared by the original dataset; please defer to the original.