Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

LIBERO Spatial, re-rendered at 224x224 (Patch Policy layout)

The 500 demonstrations of LIBERO Spatial (10 tasks x 50) from yifengzhu-hf/LIBERO-datasets (Apache-2.0), converted to the per-demo layout of the Patch Policy LIBERO Goal release, so that datasets.libero.LiberoGoalDataset(data_directory, suite="libero_spatial") reads it.

What changed from the source

  • Each demo's saved MuJoCo states are replayed in the LIBERO simulator and both cameras are rendered again at 224x224 (the source stores 128x128), flipped upright. Actions and states are copied unchanged.
  • Fixed fixtures (stove, cabinet) are welded to the world, so their placement is not in the saved states. As in the Patch Policy LIBERO Goal release, each task is replayed in one scene reset once (--seed 0), so all 50 demos of a task share one fixture placement. It is not the placement each demo was recorded in: frames where a bowl rests on or is gripped against the stove or cabinet are 1-2 cm off.
  • Made by scripts/convert_libero.py (patch_policy_cleanup, branch feat/libero-spatial). Run on LIBERO Goal, the same converter reproduces the Patch Policy release's arrays exactly and its frames up to fixture placement.

Layout

libero_spatial/<task>/demo_<k>/
    agentview_image.pth            (T, 224, 224, 3) uint8
    robot0_eye_in_hand_image.pth   (T, 224, 224, 3) uint8
    actions.npy                    (T, 7)   float64
    states.npy                     (T, 92)  float64   MuJoCo states from the source HDF5
    robot0_joint_pos.npy           (T, 7)   float64
    robot0_eef.npy                 (T, 7)   float64   position + quaternion
    robot0_gripper_pos.npy         (T, 2)   float64
    object_states.npy              (T, 70)  float64
  • 10 tasks x 50 demos, 62,250 frames, 18 GB. .pth files are PyTorch pickles: load them only if you trust this repo.

Use

hf download ParsaSharifi/patch-policy-libero-spatial --repo-type dataset --local-dir <dataset_root>/libero_dataset

then data_directory=<dataset_root>/libero_dataset, suite=libero_spatial.

Downloads last month
5

Models trained or fine-tuned on ParsaSharifi/patch-policy-libero-spatial