Installation#

RPent runs on top of a forked branch of RLinf, which supplies the LIBERO / RoboCasa simulators and the VLA training / inference stack. The two repositories are meant to live side by side in a single workspace/ directory.

Prerequisites#

  • Linux with an NVIDIA GPU (LIBERO / RoboCasa render on EGL).

  • CUDA 12.x drivers matching your GPU.

  • Python 3.10+.

  • uv (used to sync RPent’s extra dependencies on top of the RLinf-provided virtualenv).

  • git, bash, and a working C toolchain for MuJoCo / robosuite.

You will also want:

  • An API key for at least one LLM provider — Anthropic, OpenAI, or an OpenAI-compatible chat endpoint — for the reasoning brain.

  • A VLA checkpoint. For LIBERO / Pi0.5 the recommended checkpoint lives at HuggingFace: rlinf-pi05-libero-130-fullshot-sft.

1. Clone RLinf and RPent side by side#

mkdir workspace && cd workspace
# RPent depends on a forked branch of RLinf; the fork will be merged
# back to RLinf main after more iterations.
git clone https://github.com/jx-qiu/RLinf -b feature/physicalagent rlinf
git clone https://github.com/RLinf/RPent rpent

The exact layout you should end up with:

workspace/
├── rlinf/     # RLinf fork (simulators + VLA infrastructure)
└── rpent/     # RPent (the agent framework — this repo)

2. Create the RLinf virtualenv (LIBERO + openpi)#

RLinf ships a single-command installer that builds a virtualenv with LIBERO + openpi (the Pi0.5 runtime) pre-wired:

cd rlinf
bash requirements/install.sh embodied \
  --env libero --model openpi --use-mirror \
  --venv ../.venv-opi-libero
cd ..
source .venv-opi-libero/bin/activate

The installer places the virtualenv one level above rlinf/ so it can be reused by RPent. Activate it before the next step.

3. Install RPent’s extra dependencies#

With the RLinf virtualenv active, sync RPent’s extras on top of it and run the LIBERO PRO/PLUS setup script:

cd rpent
uv sync --active --inexact
bash scripts/install_libero_pro_plus.sh

uv sync --active --inexact adds RPent’s own Python dependencies (pydantic-ai, the Claude Agent SDK, the Codex SDK, FastAPI for the dashboard, …) without disturbing the RLinf packages that share the environment. install_libero_pro_plus.sh patches the LIBERO distribution to expose the pro and plus task variants used throughout RPent.

4. (Optional) RoboCasa#

RoboCasa (kitchen-scale, long-horizon manipulation) has its own one-shot setup script:

bash scripts/setup_robocasa.sh

See docs/SETUP_ROBOCASA.zh.md for the full RoboCasa365 + RLDX-1 walkthrough (assets, controller configs, VLA checkpoints).

5. (Optional) Real-world robot dependencies#

Franka and SO-101 support is being rolled in; when it lands, each robot’s driver ships as a package under robots/<name>/ with its own README.md describing the SDK / firmware requirements. See Franka and SO-101 for the current status.

Verifying the install#

The quickest way to confirm everything is wired correctly is to run one LIBERO task end-to-end — see Quick start. If that succeeds, the env server, VLA server, and reasoning brain are all healthy.

If something breaks:

  • The env server writes its stdout / stderr to <output_dir>/env_server.log.

  • The VLA server writes to <output_dir>/vla_server.log.

  • The agent’s own run log lives at <output_dir>/run.log.

The three logs are always in that per-run scratch directory, so a failed run is self-contained and easy to inspect.