Installation#

Install RLinf in one of two ways: Option 1 (UV) builds a local virtual environment matched to your machine, while Option 2 (Docker) gives the most reproducible setup.

Option 1: UV#

Build a UV virtual environment with install.sh. Pass an install target (embodied, agentic, or docs) plus any target-specific flags:

git clone https://github.com/RLinf/RLinf.git
cd RLinf
bash requirements/install.sh embodied --model openvla --env maniskill_libero
source .venv/bin/activate

More targets and combinations:

bash requirements/install.sh embodied --model openvla-oft --env maniskill_libero
bash requirements/install.sh embodied --model openpi --env libero
bash requirements/install.sh agentic
bash requirements/install.sh docs

Run bash requirements/install.sh --help for the complete model and environment list.

  • Use --venv <dir> to choose the virtual environment directory.

  • Use --use-mirror for faster downloads from mainland China.

  • Use --python <version> only when a package requires it. The default is Python 3.11.14; some environments such as behavior and d4rl require Python 3.10.

  • Use --torch <version> only when you need a different PyTorch wheel.

  • Use --platform amd or --platform ascend for experimental non-NVIDIA installs. See RL with LIBERO on AMD ROCm and RL with LIBERO on Ascend CANN.

Option 2: Docker#

Each image bundles a ready-to-run stack (for example, agentic-rlinf0.2-maniskill_libero). Pull and run it, then select the model environment inside the container:

docker pull rlinf/rlinf:agentic-rlinf0.2-maniskill_libero
docker run -it --gpus all \
   --shm-size 100g \
   --net=host \
   --name rlinf \
   -e NVIDIA_DRIVER_CAPABILITIES=all \
   rlinf/rlinf:agentic-rlinf0.2-maniskill_libero /bin/bash

git clone https://github.com/RLinf/RLinf.git
cd RLinf
source switch_env openvla

  • Keep -e NVIDIA_DRIVER_CAPABILITIES=all for GPU rendering.

  • Do not mount over /root or /opt; those directories contain assets and virtual environments in the image.

  • If your platform changes $HOME or remounts /root, run link_assets inside the container before launching an example.

  • Switch model environments with source switch_env openvla, source switch_env openvla-oft, or source switch_env openpi.

Verify#

After activation, verify that RLinf and Ray are visible in the environment:

python -c "import rlinf; print(rlinf.__file__)"
ray --version

Next Steps#