libero-r6

RGB observations, robot state and actions for the four LIBERO suites (libero_spatial, libero_object, libero_goal, libero_10), with end-effector rotations stored as 6-D rotation representations ("r6"). This is the fine-tuning data for the PointmapVLA LIBERO and LIBERO-Plus results, used together with DAVIAN-Robotics/libero-r6-pointmap which provides the matching depth and camera parameters.

• 1,693 episodes over 40 tasks (432 / 454 / 428 / 379 per suite), 20 fps
• 2 cameras at 256×256: image (agentview) and wrist_image
• Converted from the IPEC-COMMUNITY/libero_*_no_noops_1.0.0_lerobot releases (LeRobot v2.1) to LeRobot v3.0, one parquet and one video file per episode

Format

Each suite is its own LeRobot v3.0 dataset directory (libero_<suite>_no_noops_1.0.0_lerobot/).

Feature	Shape	Meaning
observation.images.image	video 256×256	agent-view RGB
observation.images.wrist_image	video 256×256	wrist RGB
observation.state	float32 (11,)	end-effector position (3), end-effector rotation as 6-D (first two rows of the rotation matrix, row-major) (6), gripper joint positions (2)
action	float32 (10,)	end-effector position delta (3), rotation delta as 6-D (6), gripper (1)

The 6-D rotation is the continuous representation of Zhou et al. (CVPR 2019), converted exactly from the source rotation vectors and axis-angle actions.

Verification

MANIFEST.json lists the size and SHA-256 of every file; check a download with python scripts/verify_data.py <local_dir> from the PointmapVLA repository.

License and attribution

Derived from LIBERO (MIT License) via the IPEC-COMMUNITY LeRobot conversions. Released under the MIT License.