THULab/demo2_frames_grab

Demo2 Frames Grab (TsFile)

Apache TsFile version of 1g0rrr/demo2_frames_grab.

Overview

This LeRobot v2.1 dataset records the task “Peel off protective film from adhesive tape.” It uses the sam_two robot type and provides numeric robot action and state observations for each frame.

• Scale: 64 train episodes, 28,432 frames, one task, sampled at 30 fps.
• Source videos: the original dataset has laptop, side, and phone camera streams. Videos are not included in this TsFile dataset; see the source videos.

Schema (TsFile structure)

• Time (INT64, milliseconds) — round(timestamp * 1000) for each source frame; time restarts for every episode.
• episode_index, task_index (TAG) — source episode and task dimensions; query a single episode with WHERE episode_index=0.
• frame_index, sample_index (FIELD, INT64) — source frame index and source index (renamed to sample_index).
• action_0 … action_6 (FIELD, FLOAT) — flattened source action[7] vector.
• observation_state_0 … observation_state_6 (FIELD, FLOAT) — flattened source observation.state[7] vector.

The source timestamp column is represented by Time and is not retained as a duplicate field. Source vectors are flattened by element index, and their order is preserved.

Usage

Read data/demo2_frames_grab.tsfile with the Apache TsFile Java or Python SDK.

Source & license

• Original dataset: https://huggingface.co/datasets/1g0rrr/demo2_frames_grab
• Author / publisher: 1g0rrr
• Source format: LeRobot v2.1
• License: Apache-2.0