Skip to content
RLinf
GitHub (opens in a new tab)

Reinforcement Learning Infrastructure for Embodied and Agentic AI

RLinf is an open-source RL infrastructure for embodied and agentic AI. Train VLA models and agents with PPO, GRPO, SAC and more, from simulation to real robots.

Get Started
Illustration of data flowing from dashboards and charts into a GPU

Runs on hardware from industry leaders:

  • NVIDIA
  • AMD
  • Huawei Ascend
  • Moore Threads
  • MetaX
  • Biren

Features

  • Embodied and Agentic RL

    Train VLA and world action models in simulation or on real robots, and agents with single- or multi-agent RL, in one framework.

  • Scale Without Code Changes

    Hide the complexity of distributed RL. Scale to many GPU nodes and switch between FSDP and Megatron backends without changing code.

  • Runs on Diverse Hardware

    One training stack runs on NVIDIA and AMD GPUs and on Huawei Ascend, Moore Threads, MetaX, Biren and more, with zero changes to models and algorithms.

Get Started in Seconds

Install RLinf with UV or Docker, then follow the guides to run your first training.

  1. Install RLinf

    Pick your setup on the right, then run the command.

  2. Run the VLA quickstart

    Train your first VLA model with the quickstart guide.

  3. Scale beyond one machine

    Launch the same training across multiple nodes.

Installation guide
Install method
Target
Model
Hardware
Mirror
Install command
git clone https://github.com/RLinf/RLinf.git
cd RLinf
bash requirements/install.sh embodied --model openvla-oft --env maniskill_libero
source .venv/bin/activate

Architecture

RLinf architecture overview

Broad Algorithm, Model & Hardware Support

Algorithms, models, simulators and hardware that RLinf works with.

RL Algorithms

See all RL Algorithms(opens in a new tab)

Models

See all Models(opens in a new tab)

Simulators

See all Simulators(opens in a new tab)

Hardware

  • NVIDIA GPUs
  • AMD GPUs
  • Huawei Ascend
  • Moore Threads
  • MetaX
  • Biren

See all Hardware(opens in a new tab)

Release Notes

v0.3

Latest218 PRs · 65 contributors · 34 new contributors

New models

  • Dexbotic DM0
  • DreamZero
  • GR00T-N1.6 / N1.7
  • ABot-M0
  • StarVLA
  • LingBot-VLA

Highlights

  • Six new embodied models, from the DreamZero world-model VLA to GR00T-N1.6 / N1.7, StarVLA and LingBot-VLA. Docs(opens in a new tab)
  • New simulators Genesis, Polaris and RoboVerse, plus LIBERO+ / LIBERO-Pro, Embodichain and RoboCasa examples. Docs(opens in a new tab)
  • An end-to-end real-world loop, from data collection to SFT, RL and deployment, with Spacemouse, VR and GELLO teleoperation. Docs(opens in a new tab)
  • New algorithms for real-world and simulation RL, including DSRL for π0.5, RECAP, SAC-Flow, DAgger and HG-DAgger. Docs(opens in a new tab)
  • End-to-end runs on Huawei Ascend (CANN) and AMD ROCm via install.sh --platform. Docs(opens in a new tab)
  • An SGLang inference server, decoupled env mode and upgraded weight synchronization. Docs(opens in a new tab)
+2 more highlights(opens in a new tab)

See all releases on GitHub(opens in a new tab)

Join the Community

From first-time users to teams training at scale, the community is open to everyone.