| # LIBERO simulation client for SN80 eval. | |
| # | |
| # Kept as a SEPARATE image from the policy server on purpose: LIBERO pins | |
| # numpy 1.22.4 / transformers 4.21 / torch 1.11, which cannot coexist with the | |
| # training image's numpy 1.26 / transformers 4.57 / torch 2.8. That is the same | |
| # split openroboto-evaluation uses - sim client here, policy over websocket. | |
| # | |
| # Adapted from openpi/examples/libero/Dockerfile, which already solves the two | |
| # things that otherwise bite: | |
| # * LIBERO prompts interactively on first import unless LIBERO_CONFIG_PATH | |
| # points at a prewritten config.yaml. | |
| # * MuJoCo needs an EGL ICD; the CUDA runtime base ships none. The actual | |
| # driver libs come from the host at run time via | |
| # NVIDIA_DRIVER_CAPABILITIES=compute,utility,graphics. | |
| FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04 | |
| COPY --from=ghcr.io/astral-sh/uv:0.5.1 /uv /uvx /bin/ | |
| RUN apt-get update && apt-get install -y --no-install-recommends \ | |
| make g++ clang \ | |
| libosmesa6-dev libgl1-mesa-glx libegl1 libglew-dev libglfw3-dev \ | |
| libgles2-mesa-dev libglib2.0-0 libsm6 libxrender1 libxext6 \ | |
| python3-dev python3.8-dev \ | |
| && rm -rf /var/lib/apt/lists/* | |
| # python3.8-dev supplies Python.h; needed to build evdev (a robosuite -> | |
| # LIBERO transitive dependency) from source, since uv's own portable Python | |
| # 3.8 ships no dev headers of its own. | |
| WORKDIR /app | |
| ENV UV_LINK_MODE=copy | |
| ENV UV_PROJECT_ENVIRONMENT=/.venv | |
| COPY requirements.txt /tmp/requirements.txt | |
| COPY requirements-libero.txt /tmp/requirements-libero.txt | |
| COPY openpi-client /tmp/openpi-client | |
| RUN uv venv --python 3.8 $UV_PROJECT_ENVIRONMENT | |
| # gym==0.25.2 (a 2022-era package) fails to build under uv's default build | |
| # isolation because uv fetches a modern setuptools whose _distutils_hack is | |
| # not Python-3.8-compatible (TypeError: 'type' object is not subscriptable | |
| # while gym's setup.py bootstraps distutils). --build-constraints pins an old | |
| # setuptools into gym's isolated build environment only, not the final venv. | |
| RUN echo "setuptools==65.5.0" > /tmp/build-constraints.txt | |
| # evdev (a robosuite -> real-hardware-teleop transitive dep, unused for | |
| # scripted LIBERO eval but still resolved and built) needs Python.h. Ubuntu | |
| # jammy ships no python3.8-dev package at all (apt silently no-ops instead of | |
| # erroring on the missing name) - but uv's OWN managed CPython 3.8.20 build | |
| # already includes its headers; the build backend just isn't finding them. | |
| # Point the compiler at them explicitly. | |
| ENV CPATH=/root/.local/share/uv/python/cpython-3.8.20-linux-x86_64-gnu/include/python3.8 | |
| RUN uv pip sync /tmp/requirements.txt /tmp/requirements-libero.txt \ | |
| /tmp/openpi-client/pyproject.toml \ | |
| --extra-index-url https://download.pytorch.org/whl/cu113 \ | |
| --index-strategy=unsafe-best-match \ | |
| --build-constraint /tmp/build-constraints.txt | |
| # LIBERO + LIBERO-PRO + the openpi client sources live under /app. | |
| COPY libero /app/third_party/libero | |
| COPY libero-pro /app/third_party/libero-pro | |
| COPY openpi-client /app/packages/openpi-client | |
| COPY main.py /app/examples/libero/main.py | |
| ENV PYTHONPATH=/app:/app/packages/openpi-client/src:/app/third_party/libero | |
| # Prewritten so LIBERO's __init__ never prompts. | |
| ENV LIBERO_CONFIG_PATH=/tmp/libero | |
| RUN mkdir -p /tmp/libero && printf '%s\n' \ | |
| 'benchmark_root: /app/third_party/libero/libero/libero' \ | |
| 'bddl_files: /app/third_party/libero/libero/libero/bddl_files' \ | |
| 'init_states: /app/third_party/libero/libero/libero/init_files' \ | |
| 'datasets: /app/third_party/libero/libero/datasets' \ | |
| 'assets: /app/third_party/libero/libero/libero/assets' \ | |
| > /tmp/libero/config.yaml | |
| # EGL ICD pointing at the driver lib below. | |
| RUN mkdir -p /usr/share/glvnd/egl_vendor.d && \ | |
| printf '%s' '{"file_format_version":"1.0.0","ICD":{"library_path":"libEGL_nvidia.so.0"}}' \ | |
| > /usr/share/glvnd/egl_vendor.d/10_nvidia.json | |
| # NVIDIA's EGL/GLX userspace libraries, vendored into the image, because this | |
| # host runs a compute-only datacenter driver install: `--gpus all` + | |
| # NVIDIA_DRIVER_CAPABILITIES=graphics mounts NOTHING for EGL here, because the | |
| # HOST ITSELF has no libEGL_nvidia.so.0 anywhere on its filesystem - confirmed | |
| # by `find / -iname libEGL_nvidia*` on the host coming back empty. Without | |
| # this, MjRenderContextOffscreen raises "Cannot initialize a EGL device | |
| # display... does not support the PLATFORM_DEVICE extension" on first | |
| # env.reset(), i.e. only once real rendering is attempted - every earlier | |
| # import succeeds, which makes this look like a rendering-logic bug rather | |
| # than a missing-library one. | |
| # | |
| # Fix: extract ONLY the userspace GL/EGL libs (not the kernel module, not the | |
| # CUDA driver, not the installer's dkms/nvidia-smi bits) from the tesla/ .run | |
| # matching this host's exact driver version, via --extract-only: | |
| # 580.126.20: https://us.download.nvidia.com/tesla/580.126.20/NVIDIA-Linux-x86_64-580.126.20.run | |
| # ./NVIDIA-Linux-x86_64-580.126.20.run --extract-only --target /tmp/x | |
| # then take libEGL_nvidia.so, libnvidia-eglcore.so, libnvidia-glcore.so, | |
| # libnvidia-glsi.so, libnvidia-tls.so, libnvidia-gpucomp.so, libGLX_nvidia.so, | |
| # libglxserver_nvidia.so. MUST match the host driver version exactly - these | |
| # are not forward/backward compatible across driver releases. Re-extract and | |
| # rebuild if the host driver is ever upgraded. | |
| COPY nvidia-egl-libs /opt/nvidia-egl | |
| ENV LD_LIBRARY_PATH=/opt/nvidia-egl | |
| ENV MUJOCO_GL=egl | |
| ENV PYOPENGL_PLATFORM=egl | |
| CMD ["/bin/bash"] | |