Datasets:

ArXiv:
License:
roboverse_data / libero /README.md
somaghr's picture
Add libero/: 130 base LIBERO tasks vendored as portable MJCF + 193 deduped shared assets (library-free; checker 130/130 bitwise)
34fdeaa verified
|
raw
history blame
2.4 kB

LIBERO — vendored assets for library-free reproduction in RoboVerse / MetaSim

This libero/ subtree of RoboVerseOrg/roboverse_data makes the upstream libero and robosuite Python packages deletable: every one of the 130 base LIBERO tasks (5 suites) can be loaded, evaluated and rendered inside MetaSim's MuJoCo substrate using only these files + mujoco+numpy (no libero, no robosuite import).

Layout

libero/
  manifest.json                       # task list + asset stats + sha
  assets/robosuite/<...>              # Franka / Panda gripper / arena meshes (shared)
  assets/libero/<...>                 # chiliocosm objects, textures, meshes
  tasks/<suite>/<task>.xml            # portable MJCF — file="…" are asset relpaths
  tasks/<suite>/<task>.goal.json      # resolved BDDL goal + OSC config + body_pos/quat
  • 193 unique assets / 265 MB, deduplicated across all 130 tasks (vs ~30 GB if each task embedded its own meshes). robosuite/ (64 files) is shared by every task; libero/ (129 files) holds the per-scene objects/textures.
  • Each <task>.xml is the live robosuite-compiled MJCF with file= paths rewritten to the two-namespace relpaths above (resolve with assets/<relpath>).
  • <task>.goal.json carries the BDDL goal reduced to MuJoCo-evaluable predicates (in/on/on_site/open/close/turnon/turnoff), the OSC_POSE config, and the captured body_pos/body_quat (robosuite's runtime fixture placement, seed=0) which the loader re-applies to make the model bitwise-identical.

How it's produced & consumed (RoboVerse)

  • Produced once by scripts/native/migrate_libero_assets.py.
  • Consumed by roboverse_pack/tasks/libero_native: LiberoNativeTask.from_vendored(suite, task) resolves the MJCF + every mesh via _locator.py (this dataset, local clone or HF download) — never importing libero.

Verification (scripts/native/verify_vendored.py)

All 130 tasks, vendored vs the original robosuite env:

  • model fields exact: 130/130 (max|Δ| = 0 after the body_pos/quat re-apply)
  • ported BDDL checker 1:1: 130/130 (0 mismatch vs env._check_success over full demo state-replay)

Render via MJCF recompile is ~2–5 MAE/255 vs robosuite (mesh-recompile shading only); physics and success-checking are bitwise.