LIBERO native-action EEF10 (LeRobot v3.0)

This local dataset was converted from lerobot/libero at revision a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 (local source: /mnt/data/wangyuran/libero-lerobot). Its original parquet packing, episode offsets, tasks, 10 FPS timeline, and videos are preserved.

Both primary robot columns use a 10-D interface:

• observation.state: achieved [xyz3, rot6d6, gripper_open_scale1]
• action: native normalized LIBERO [delta_xyz3, delta_rot6d6, gripper_open_command1]

The action xyz values are copied unchanged from the source. The action rotation is rot6d(Exp(source_delta_axis_angle)); no 0.05 m or 0.5 rad OSC scale is applied and it is not composed with the achieved state. Native LIBERO uses +1 = close; the action gripper is negated to match the OpenWAM pretraining convention -1 = closed, +1 = open. State gripper is projected from the achieved finger separation as clip(2 * (qpos0 - qpos1) / 0.08 - 1, -1, 1).

The training/deployment normalization artifact is meta/normalization_stats.npy. It contains separate libero (action) and libero_state (state) blocks, and no pooled compatibility block. The canonical LIBERO reader uses the separate blocks so achieved-state and command statistics are never mixed. Its six rot6d dimensions are hard identity-mapped during normalization, even if a stale/custom stats file is supplied.

Dataset totals: 1693 episodes, 273465 frames, 40 tasks, 10.0 FPS. Full provenance and numerical round-trip checks are recorded in meta/conversion.json.