Myungkyu/rmbench_lerobot

RMBench — LeRobot (GR00T-compatible), 9 official tasks

A GR00T-compatible [LeRobot](https://github.com/huggingface/lerobot) v2.1 conversion of RMBench, a dual-arm robotic memory manipulation benchmark. 450 episodes (9 tasks × 50 expert demos), 277,350 frames, 3 camera views, 14-D absolute-joint action.

Source

Derived from the official RMBench release [`TianxingChen/RMBench`](https://huggingface.co/datasets/TianxingChen/RMBench) (demo_clean split): RoboTwin-2.0 simulator, dual-arm Aloha-AgileX (6-DoF arm + 1 gripper per arm), per-episode HDF5, 50 expert demos/task. License MIT (inherited).

What was converted

A format / representation conversion only — the underlying trajectories are unchanged from the official release.

• —Tasks — kept only the 9 tasks in the RMBench paper: observe_and_pickup, rearrange_blocks, put_back_block, swap_blocks, swap_T, battery_try, blocks_ranking_try, cover_blocks, press_button.
• —Action / state (14-D) — reordered to contiguous `[left_arm(6), right_arm(6), left_gripper, right_gripper]`, absolute joint positions; action[t] = state[t+1].
• —Cameras → 3 views — head→front_view, left→left_wrist_view, right→right_wrist_view.
• —Video — re-encoded to h264, stored in videos/ (not embedded in the parquet files).
• —Format — LeRobot v2.1 with meta/{info,modality,stats}.json, ready for the GR00T data loader.

Layout

data/    # per-episode parquet (proprioceptive state + action)
videos/  # h264 video, 3 views per episode
meta/    # info.json, modality.json, stats.json

Usage (GR00T N1.6 ± HAMLET)

huggingface-cli download --repo-type dataset Myungkyu/rmbench_lerobot --local-dir data/rmbench
# then fine-tune with the public repo (HAMLET-Isaac-GR00T-N1.6):
DATASET_PATH=data/rmbench bash run_scripts/rmbench/train_hamlet_n1d6.sh

Citation

This dataset is a format conversion only; please cite the original RMBench and RoboTwin works (and follow the source dataset's MIT license).