dinyu/libero-spatial-rot

LIBERO-Spatial, multi-angle agentview

LIBERO-Spatial re-rendered with the agentview camera rotated around the robot base, at 9 angles: 0°, 30°, 60°, 90°, 120°, 240°, 270°, 300°, 330°.

The point is viewpoint generalization. Policies trained on LIBERO see exactly one agentview pose, and they tend to fall apart when that camera moves. This dataset holds the demonstrations fixed and varies only the camera, so a model can be trained across viewpoints — and, because the 150°/180°/210° band is deliberately absent, tested on angles it never saw.

Built with libero_rot/.

Contents

Format	LeRobot v3.0
Angles	9 (theta000 … theta330, 30° steps, 150/180/210 omitted)
Tasks	10 (all of LIBERO-Spatial)
Episodes	490 per angle (49 per task) — 4,410 total
Frames	60,974 per angle — 548,766 total
fps	20
Robot	Franka Panda
Size	~19 GB

Every angle contains the same 490 demonstrations, so the 9 subsets are aligned episode-for-episode. They differ only in the observation.images.image video.

libero_spatial_theta000_no_noops_lerobot/
libero_spatial_theta030_no_noops_lerobot/
...
libero_spatial_theta330_no_noops_lerobot/
    data/chunk-000/file-000.parquet
    videos/observation.images.image/chunk-000/file-000.mp4
    videos/observation.images.wrist_image/chunk-000/file-000.mp4
    meta/{info.json, tasks.parquet, stats.json, episodes/}
text_embeds_cache/          # precomputed T5 embeddings for the 10 task strings

Features

key	shape	dtype
observation.images.image	512×512×3	video (agentview, rotated)
observation.images.wrist_image	512×512×3	video (eye-in-hand)
observation.state	8	float32
observation.states.ee_state	6	float32
observation.states.joint_state	7	float32
observation.states.gripper_state	2	float32
action	7	float32

The wrist camera is rigidly attached to the end effector, so it is unaffected by the agentview rotation. It is rendered once per demonstration and shared across all 9 angles.

⚠️ Image orientation — read this before evaluating

LIBERO's raw obs images are vertically flipped. There are two conventions in circulation, and mixing them up will silently wreck your evaluation:

convention	transform	used by
[::-1, ::-1] (180° rotation / mirrored)	vertical + horizontal flip	openvla modified_libero_rlds, and the FastWAM public LIBERO datasets derived from it
[::-1] (vertical flip only)	vertical flip	this dataset

This dataset uses the visually-correct, un-mirrored orientation. With FastWAM, that means you must set:

export FASTWAM_LIBERO_IMAGE_FLIP=v      # NOT the default "vh"

Camera rotation

agentview is rotated about the z-axis of robot0_link0 (the robot base) by theta degrees. In LIBERO the base sits at the origin with identity orientation, so this is equivalent to a rotation about the world z-axis.

def rotated_camera_pose(cam_pos, cam_quat, base_pos, base_quat, theta_deg):
    cam_world     = Pose(cam_pos, cam_quat)
    base_world    = Pose(base_pos, base_quat)
    world_to_base = Pose().transform(base_world.inverse())
    cam_base      = world_to_base.transform(cam_world)
    rot           = Pose(np.zeros(3), euler.euler2quat(0, 0, np.deg2rad(theta_deg)))
    new_cam_base  = rot.transform(cam_base)
    return world_to_base.inverse().transform(new_cam_base)

theta000 is the stock LIBERO agentview pose.

How it was built

Two stages, both in libero_rot/.

Stage A — re-render. Each demo's states are restored frame by frame with sim.set_state_from_flattened and re-rendered at every angle. This replays simulator state rather than actions, so there is no rollout drift and the 9 angles stay exactly in sync.

Stage B — pack. One LeRobot v3.0 dataset per angle, matching the schema of yuanty/LIBERO-fastwam. Per-episode mp4s are concatenated losslessly with the ffmpeg concat demuxer, with from_timestamp/to_timestamp recorded in the episode metadata.

Filtering, matching openvla's modified_libero_rlds:

• no-op frames dropped (‖action[:-1]‖ < 1e-4 and gripper unchanged)
• demonstrations whose final state fails check_success() are excluded

That filtering is why there are 490 episodes rather than the raw 500.

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("dinyu/libero-spatial-rot",
                    root="libero_spatial_theta060_no_noops_lerobot")

To train across viewpoints, point your dataset config at several angle directories at once — see configs/data/libero_spatial_rot9_lerobot_v30.yaml in the companion repo.

Held-out angles

150°, 180° and 210° were never rendered. They are the generalization test: LIBERO can be evaluated at an arbitrary angle at runtime, so you can measure success at angles absent from training and compare against the trained ones.

export FASTWAM_LIBERO_IMAGE_FLIP=v
export FASTWAM_LIBERO_CAMERA_THETA=180

Citation

The underlying demonstrations are from LIBERO. Please cite the original work:

@article{liu2023libero,
  title   = {LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author  = {Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and
             Liu, Qiang and Zhu, Yuke and Stone, Peter},
  journal = {arXiv preprint arXiv:2306.03310},
  year    = {2023}
}