LIBERO-Plus SIM server

Run a pinned LIBERO-Plus simulator independently of the model/evaluator Python environment. This benchmark owns native tasks, perturbations, observations, actions and success. Policies stay in AgentPolicies/<profile>; shared transport and records stay in harness.

Evaluation protocol and 10,030-variant inventory
Server parameters
Shared server interface and evaluator launch
Retired StarVLA model launchers remain in Git history, not the active interface

Directory responsibilities

server.py is the public factory/launch entrypoint. bench_server/adapter.py owns the public sensor/action interface and native score bridge; bench_server/protocol/catalog.py and catalogs/ define the immutable task matrix. Native installation, source pins and catalog-generation tools belong in environment/; in particular native_source_manifest.json is a runtime pin, not a task catalog. checks/ contains offline regressions. There is no old root adapter/catalog shim. Relocation preserves catalog bytes and native behavior; operator paths and selected cases stay in the server configuration.

Configure a native environment

Use a separate Python 3.10 environment; Plus and standard LIBERO both export the libero namespace and must not coexist in one running process. Do not change someone else's native checkout, environment or ~/.libero/config.yaml.

Clone the official repository into your own new directory and check out commit 4976dc30028e805ff8094b55501d532c48fec182.
Download assets.zip from the pinned official public release. The archive is 6,395,849,578 bytes with SHA-256 96764a4bfbdaea98d4411598caeab235458318fe0f549611b93d1a323027b3cf. Use the checked extraction command below: this release retains a long upstream build-path prefix, so a plain unzip into the checkout creates the wrong directory layout. The tool strips only the exact reviewed prefix. It requires a new asset directory and never merges or removes prior data. BDDL and init files come from the pinned source checkout. The ZIP has 457,675 entries and about 9 GB of uncompressed file contents; allow at least 20 GB plus filesystem inode headroom for archive and extracted storage. Training datasets and model weights are not needed.
Install native EGL/OpenGL and ImageMagick Wand support. On a compatible Ubuntu NVIDIA server this normally includes libegl1, libgl1, libexpat1, libfontconfig1, and libmagickwand-dev. Host NVIDIA drivers and a working EGL vendor library are still required; pip alone cannot install the host driver. Use an isolated container/environment when system package changes are inappropriate.
In the new Python environment, install Torch 2.5.1+cpu from the official CPU wheel index, then runtime-requirements.txt, then the pinned native checkout with pip install --no-deps -e <checkout>. The native loader relies on Torch's pre-2.6 torch.load behavior; silently substituting a newer Torch is not this track. Configure the RSI checkout on PYTHONPATH or install RSI editable without replacing native dependencies.

From the RSI source root, with placeholder paths replaced:

python -B -m benchmarks.simBenchmarks.LIBERO-plus.environment.extract_assets \
  --archive /absolute/path/to/downloaded/assets.zip \
  --destination /absolute/path/to/LIBERO-plus/libero/libero/assets \
  --receipt /absolute/path/to/private-install-records/assets.json
python -B -m benchmarks.simBenchmarks.LIBERO-plus.environment.configure \
  --upstream /absolute/path/to/LIBERO-plus \
  --config-dir /absolute/path/to/libero-plus-runtime-config
export LIBERO_CONFIG_PATH=/absolute/path/to/libero-plus-runtime-config
export MUJOCO_GL=egl
export PYOPENGL_PLATFORM=egl

The extractor validates release identity, paths, entry types, duplicates and expansion bounds before creating the target. Extraction failures preserve the partial directory and do not silently restart. Hundreds of thousands of tiny asset files can be very slow on shared storage; prefer an owned node-local deployment when available. A tmpfs deployment is ephemeral: retain the verified archive and source pin on persistent storage and rebuild after a node restart.

The configuration tool only creates an explicit external path config and refuses to overwrite a different one. Catalog and source-pin verification require no reset, rendering, model, training data or policy API key.

python -B -m benchmarks.simBenchmarks.LIBERO-plus.environment.verify_environment \
  --upstream /absolute/path/to/LIBERO-plus --check-imports

This checks the 60 declared source files, resolved BDDL/reset bytes for all 10,030 variants, required asset-folder presence and native imports. It does not deserialize reset pickles, construct a simulator, validate all asset contents, or prove that EGL renders valid images. Unrecorded Python, compiled extensions and bytecode caches in the executed native package are rejected, including an added outer namespace __init__.py. Start with -B and a clean owned clone; do not remove caches from someone else's tree.

Start a fixed SIM service

Copy server.example.json to a private configuration file, replace the native path and preselect the task IDs, seed and budgets before running any model. Set RSI_SIM_SERVER_TOKEN using your evaluator's private secret provisioning; never put its value into a policy prompt, tracked JSON or shell argument.

python -B -m benchmarks.simBenchmarks.LIBERO-plus.server \
  --parameters /absolute/path/to/libero-plus-server.json \
  --seed 20260907 --episodes 1 --port 8765 \
  --output-dir AgentPolicies/<profile>/results/libero_plus/raw/<task-description>/server

The service binds only 127.0.0.1 and initializes no simulator until the trusted evaluator opens/resets the selected session. Use the neutral runner's --benchmark libero_plus --sim-server http://127.0.0.1:8765 from a separate evaluator environment, following the shared service guide. Cross-node use needs an SSH tunnel, not a publicly exposed server. Every selected session is claimed once, including errors; do not retry a failed control acknowledgment.

Validate without interpreting it as a model score

CPU contract checks:

python -B -m unittest discover \
  -s benchmarks/simBenchmarks/LIBERO-plus/checks -p 'test_*.py'

Before capability evaluation, explicitly select a separate fixed-action infrastructure diagnostic through the same SIM service. Retain reset/frame/action receipts and check independent camera views, orientation stability and native control counts. This does not award a model score or authorize repeating a previous capability attempt. Robosuite 1.4's native renderer lifecycle requires particular care after hard resets; a working EGL import is not proof that all post-reset images are valid. No image-repair or renderer monkeypatch is applied silently by this adapter.