clarify libero wrapper docs

Author: JJJYmmm

eval/libero/README.md

@@ -151,10 +151,24 @@ CMAKE_CUDA_ARCHITECTURES=80 \
 bash eval/libero/run_model_server.sh
 ```
 
-Common overrides are environment variables: `BUILD_DIR`, `ROBOTCPP_BACKEND`,
-`HOST`, `PORT`, `SUITE`, `TASK_IDS`, `N_EPISODES`, `SEED`, `EPISODE_LENGTH`,
-`MUJOCO_GL`, `PYOPENGL_PLATFORM`, and `OUTPUT`. Extra shell arguments are passed
-to `python -m eval.libero.run_model_server` before `--server-command`.
+Configure the wrapper with environment variables:
+
+| Variable | Purpose |
+| --- | --- |
+| `BUILD_DIR`, `ROBOTCPP_BACKEND`, `CMAKE_CUDA_ARCHITECTURES` | Build and backend selection. |
+| `GGUF_DIR`, `MODEL`, `SERVER_BIN` | Split GGUF inputs and `model-server` binary. |
+| `HOST`, `PORT` | Shared client/server endpoint. Set these here so both sides match. |
+| `SUITE`, `TASK_IDS`, `N_EPISODES`, `SEED`, `EPISODE_LENGTH` | LIBERO rollout selection. |
+| `MUJOCO_GL`, `PYOPENGL_PLATFORM`, `OUTPUT` | Rendering and result output. |
+
+Arguments after `run_model_server.sh` are passed to
+`python -m eval.libero.run_model_server` before the generated
+`--server-command` block. Use them for runner-only flags:
+
+```sh
+OUTPUT=eval/results/pi0-libero-object.json \
+bash eval/libero/run_model_server.sh --episode-length 400
+```
 
 Run the Python rollout runner directly when the server command should be fully
 explicit:

eval/libero/run_model_server.sh

@@ -5,20 +5,21 @@ usage() {
     cat <<'EOF'
 Run LIBERO eval against a freshly built model-server.
 
-Common overrides:
-  CONDA_ENV=vlacpp-libero   optional conda env for the Python eval runner
-  ROBOTCPP_BACKEND=cuda     model-server backend: cuda or cpu
-  BUILD_DIR=build-cuda      CMake build directory; defaults to build for cpu
-  GGUF_DIR=...              split GGUF checkpoint directory
-  MODEL=...                 split GGUF filename prefix
-  SUITE=libero_object       LIBERO suite
-  TASK_IDS=0                comma list of task ids
-  N_EPISODES=1              episodes per task
-  CMAKE_CUDA_ARCHITECTURES=80
-
-Extra arguments are passed to eval.libero.run_model_server before the
-model-server command. Configure host/port through HOST and PORT so the eval
-client and launched server stay in sync.
+Wrapper configuration:
+  CONDA_ENV=vlacpp-libero      optional conda env for the Python eval runner
+  ROBOTCPP_BACKEND=cuda        model-server backend: cuda or cpu
+  BUILD_DIR=build-cuda         CMake build directory; defaults to build for cuda
+  GGUF_DIR=...                 split GGUF checkpoint directory
+  MODEL=...                    split GGUF filename prefix
+  HOST=127.0.0.1 PORT=5555     shared client/server endpoint
+  SUITE=libero_object          LIBERO suite
+  TASK_IDS=0 N_EPISODES=1      rollout selection
+  OUTPUT=...                   optional result JSON path
+  CMAKE_CUDA_ARCHITECTURES=80  optional CUDA architecture override
+
+Arguments after this script are passed to eval.libero.run_model_server before
+the generated --server-command block. Use HOST and PORT env vars instead of
+extra --host/--port flags so the eval client and launched server stay in sync.
 EOF
 }