BeingBeyond/Being-H-EDU_SO101

Being-H-EDU SO101 Pick Cube Plate Datasets

This directory contains LeRobot-format SO101 datasets used by Being-H-EDU for the pick_cube_plate manipulation task.

Task

The task instruction is:

Pick the cube into the plate.

In each episode, an SO101 follower arm observes a tabletop scene and executes a pick-and-place behavior: grasp the cube, move it above the plate, and release it inside the plate. The datasets can be used with the Being-H-EDU tutorial workspace to post-train Being-H0.5-based policies, but they are not tied to Being-H-EDU only. They are standard LeRobot-style robot datasets and can also be loaded by other imitation-learning or robot-policy training pipelines that support the same schema.

Dataset Variants

Directory	LeRobot version	Episodes	Frames	Intended use
pick_cube_plate	v2.1	210	51,987	Raw merged dataset before quality filtering. Use for auditing or if you want to apply your own filtering.
pick_cube_plate_filtered	v2.1	189	45,001	Quality-filtered dataset. Short or likely incomplete episodes were removed.
pick_cube_plate_trimmed	v2.1	189	35,006	Recommended v2.1 training set. It keeps the filtered episodes and trims static or uninformative parts.
pick_cube_plate_v3.0	v3.0	210	51,987	Same raw task data packaged in the LeRobot v3.0 layout for newer LeRobot tooling.

Suffix Meaning

• No suffix (pick_cube_plate): raw v2.1 export after merging the collection sessions. It preserves all 210 episodes.
• filtered: episodes flagged by the filtering report were removed. In this release, 21 short episodes were removed from the raw 210 episodes.
• trimmed: starts from the filtered set, then removes extra static frames such as long idle segments before or after the actual manipulation. This reduces the frame count while keeping the same 189 successful/usable episodes.
• v3.0: LeRobot v3.0 storage layout. Choose this if your loader expects v3.0-style consolidated parquet/video files.

Schema

All variants describe the same SO101 task and expose the same core features:

• observation.state: 6-D robot joint/gripper state.
• action: 6-D robot joint/gripper action.
• observation.images.wrist: wrist camera video, 240 x 320 RGB, 30 FPS.
• observation.images.external: external camera video, 480 x 640 RGB, 30 FPS.
• task_index: language task index, mapped to "Pick the cube into the plate."

The robot type recorded in meta/info.json is so_follower, and the dataset FPS is 30.

Which One Should I Use?

For Being-H-EDU post-training, start with pick_cube_plate_trimmed unless you specifically need the raw data. It removes low-quality episodes and unnecessary static frames, so training focuses more on the manipulation behavior.

Use pick_cube_plate_filtered if your training pipeline benefits from keeping the full temporal context of each successful episode but you still want the short/incomplete episodes removed.

Use pick_cube_plate or pick_cube_plate_v3.0 if you want to inspect the original trajectories, reproduce preprocessing, or run your own filtering and trimming.

Using With Being-H-EDU Tutorials

Register the local dataset path through dataset_path_overrides:

so101_posttrain:
  dataset_names:
  - so101.pick_cube_plate

  dataset_path_overrides:
    so101.pick_cube_plate: /path/to/pick_cube_plate_trimmed

  data_config_names:
  - "so101"
  embodiment_tags:
  - "so101"

Being-H-EDU expects the first five SO101 action dimensions to be deltas relative to observation.state, while the gripper action remains absolute. After downloading the dataset to a writable local copy, run the conversion script from the Being-H-EDU tutorial workspace before training:

python examples/so101/convert_so101_actions_to_delta.py /path/to/pick_cube_plate_trimmed

See tutorials/Being-H-EDU/docs/so101_data_processing.md and tutorials/Being-H-EDU/docs/so101_quickstart.md in the Being-H repository for the full training workflow.