Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
180
4.1k
label
class label
190 classes
000
101
101
101
202
202
202
303
405
405
405
405
506
506
607
708
708
809
809
90a
100b
100b
100b
110c
120d
130e
1410
1410
1410
1512
1613
1714
1817
1817
1817
1918
1918
201b
201b
211c
211c
221d
221d
231e
241f
2520
2520
2624
2624
2624
2624
2726
2827
2928
2928
2928
3029
3029
312a
322b
332d
342f
3530
3530
3631
3732
3732
3732
3833
3833
3934
3934
3934
4035
4137
4238
4238
4339
443b
453c
453c
463f
4741
4844
4945
5047
5148
5148
5249
534b
534b
534b
544e
544e
544e
544e
554f
554f
554f
554f
End of preview. Expand in Data Studio

LIBERO EDA Hosted Data

Immutable, browser-oriented data for the LIBERO EDA web application.

This repository contains:

  • metadata for 130 Original LIBERO task families;
  • browser replay metadata for 6,500 Original LIBERO demonstrations and 14,347 LIBERO-Plus training trajectories;
  • gzip-compressed Arrow IPC trajectory series;
  • 128 x 128 WebP thumbnails;
  • derived Original LIBERO MP4 and MuJoCo GLB assets;
  • content-addressed LIBERO-Plus evaluation geometry and source textures, plus numeric initial-state scene descriptors for all 10,030 evaluation conditions.

LIBERO-Plus training video is not duplicated. Replay manifests link to the pinned Sylvest/libero_plus_lerobot revision. LIBERO-Plus evaluation classification, BDDL, and init-state files are not copied; the application loads definitions directly from sylvestf/LIBERO-plus@4976dc30028e805ff8094b55501d532c48fec182. The distributed scene descriptors were derived with official state index 0, five zero actions, and a recorded environment/reset seed of 10,000. They retain source-relative paths and SHA-256 provenance. If robosuite rejects a transient constructor placement, the exporter advances the same seeded random stream without reseeding, matching LIBERO's reset rejection sampling, and stops after 100 attempts. The scene manifest also pins the simulator asset dataset to Sylvest/LIBERO-plus@dd2bd61b7d9a6fef1abc52d606e983b41886a149, records the source archive SHA-256, and requires the exact 448,799-file extraction before export. The extraction is matched to the official archive by path, byte size, CRC, and a deterministic content tree SHA-256. Symlinks and incomplete or modified extracts are rejected. The pinned tree digest (libero-plus-asset-tree-sha256/v1) is 6c4c2e638f6401304f01b2573c80af41b35b6d94838df71f6ab91f59468b7ecb.

Original LIBERO video is displayed in its recorded source orientation. The published LIBERO-Plus MP4 convention is rotated 180 degrees from raw LIBERO; each replay manifest declares that convention and the Replay UI restores the simulator orientation by default. The source pixels are not rewritten. The public LIBERO-Plus files are one MP4 per episode, so hosted v2 manifests use a zero-based interval from 0 to state_count / 20. The one-way v1-to-v2 migration verifies the older chunk-global interval length before normalizing it; video bytes are neither copied nor modified.

Evaluation scenes are interactive initial states, not screenshots, videos, or successful trajectories. Geometry is shared per Original LIBERO source task, and textures are deduplicated by content hash.

manifest.json is the small browser entry point for schema, counts, and provenance. It pins integrity/artifacts.json, which records every distributed file's byte size and SHA-256 value. Both are written only after a complete export.

License and attribution

This is a mixed-license dataset. Do not treat the repository as uniformly Apache-2.0. See DATA_LICENSES.md before redistribution.

The Original LIBERO demonstration derivative retains the source attribution and CC BY 4.0 requirements. LIBERO-Plus material retains its upstream MIT license. LIBERO EDA's original catalog/export code is Apache-2.0.

Downloads last month
5,517