feat(envs): add LIBERO-plus robustness benchmark

[pkooij]

Title

feat(envs): add LIBERO-plus robustness benchmark

Type / Scope

• Type: Feature
• Scope: src/lerobot/envs/, pyproject.toml, docker/, docs/

Summary / Motivation

LIBERO-plus is a robustness benchmark for VLA models that extends LIBERO with 7 perturbation dimensions (camera viewpoints, object layouts, robot initial states, language instructions, lighting, background textures, sensor noise), producing ~10 000 task variants across the standard LIBERO suites.

Because LIBERO-plus keeps the same Python gym interface as the original LIBERO, the integration is minimal: a one-line config subclass, an import fallback for the different package nesting, and a new pip extras group.

Related issues

• Related: LIBERO integration (already in feat/async-vector-env)

What changed

• src/lerobot/envs/libero.py — wraps the top-level LIBERO imports in try/except to handle the extra module nesting level that LIBERO-plus ships with
• src/lerobot/envs/configs.py — adds LiberoPlusEnv config (@EnvConfig.register_subclass("libero_plus")), a thin subclass of LiberoEnv with task="libero_spatial" as default; fully inherits create_envs and get_env_processors
• pyproject.toml — adds libero_plus optional dep group and includes it in all
• docs/source/libero_plus.mdx — new benchmark doc: perturbation dimensions, task suites, install instructions, eval commands, camera name mapping, dataset reference
• docs/source/_toctree.yml — registers new doc page
• docker/Dockerfile.benchmark.libero_plus — isolated CI image (adds libexpat1 libfontconfig1-dev libmagickwand-dev system deps required by LIBERO-plus)
• .github/workflows/benchmark_tests.yml — adds libero-plus-integration-test job (build image + 1-episode smoke eval on aws-g6-4xlarge-plus)

No breaking changes. env.type=libero continues to work unchanged.

Dataset note

pepijn223/libero_plus_lerobot is already LeRobot v3.0 format — no conversion needed.
Dataset card (README) is missing on the Hub and should be added in a follow-up.
Camera keys: observation.images.front / observation.images.wrist.

How was this tested (or how to run locally)

• pre-commit run -a passes on all changed files
• Registration verified locally via PYTHONPATH override:

  from lerobot.envs.configs import EnvConfig
  cfg = EnvConfig.get_choice_class('libero_plus')()
  # → type=libero_plus, task=libero_spatial
  

• Full eval smoke-test (requires Linux + GPU + LIBERO-plus installed):

  lerobot-eval \
    --policy.path=pepijn223/smolvla_libero \
    --env.type=libero_plus \
    --env.task=libero_spatial \
    --eval.batch_size=1 --eval.n_episodes=1 \
    --eval.use_async_envs=false --policy.device=cuda \
    '--env.camera_name_mapping={"agentview_image": "camera1", "robot0_eye_in_hand_image": "camera2"}' \
    --policy.empty_cameras=1

  Runs automatically via the new CI job on aws-g6-4xlarge-plus.

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest) — LIBERO-plus requires Linux, validated via CI
• Documentation updated (docs/source/libero_plus.mdx)
• CI is green

Reviewer notes

• The import fallback in libero.py is the only change touching existing code paths. The try branch runs for hf-libero; the except branch for LIBERO-plus. Transparent to callers.
• LiberoPlusEnv is intentionally a minimal subclass — no duplicated logic.
• The Docker image uses uv sync --extra libero_plus --no-cache (no --locked) because the GitHub-sourced package is not in uv.lock. Pin a commit SHA in the dep once LIBERO-plus stabilizes.