typoverflow/libero_spatial_mj332

libero_spatial_no_noops_lerobot: detailed LeRobot v3.0

This dataset was converted from the Fast-WAM LIBERO MuJoCo 3.3.2 LeRobot v2.1 libero_spatial_no_noops_lerobot partition. It contains successful demonstrations whose historical no-op actions were removed by simulator replay before this conversion. This converter preserves all remaining frames and does not apply any additional filtering.

The original 8D state and 7D action vectors are preserved exactly as raw_state.ref_state and raw_action.ref_action. Canonical low-dimensional fields follow failure_rollout_data/dataset.md; debug.gripper_eef_* contains ground-truth next-step relative EEF motion for inspection. Source joint positions are exposed as raw_state.joint_pos and state.joint_pos.

Required camera transform for canonical training

The source observation.images.image and observation.images.wrist_image videos are preserved unchanged. For canonical training, horizontally flip both camera views at load time. The videos are deliberately not rewritten or re-encoded.

The source replay pipeline vertically flips the raw robosuite render and then rotates it by 180 degrees. Those vertical components cancel, leaving the stored image horizontally mirrored.

# NumPy (..., height, width, channels)
image = np.flip(image, axis=-2)

# PyTorch (..., channels, height, width)
image = torch.flip(image, dims=(-1,))

Apply this only to camera pixels. Do not flip or negate any low-dimensional field.

Conversion notes

• The source action gripper is already binary: 0=closed, 1=open.
• No frames were filtered during this conversion; see meta/noop_audit.json.
• Rotation and reconstruction checks are in meta/conversion_validation.json.
• Controller, alignment, filtering, camera, and source assumptions are in meta/conversion_config.json.
• Every numeric and video statistic includes q01 and q99.