RLinf/RLinf

RLinf: Reinforcement Learning Infrastructure for Embodied and Agentic AI

What it solves

RLinf provides a scalable and flexible infrastructure for training reinforcement learning (RL) models, specifically targeting Embodied AI (robotics) and Agentic AI. It removes the complexity of distributed programming, allowing developers to scale training across many GPU nodes without changing their code, and optimizes throughput for high-performance RL training.

How it works

RLinf acts as a robust backbone for RL training, supporting various algorithms (such as PPO, GRPO, and SAC) and integrating with multiple backends like FSDP, HuggingFace, SGLang, and vLLM. It employs a hybrid execution mode for embodied RL to increase throughput and supports a wide range of accelerators including NVIDIA GPUs, Moore Threads (MUSA), Huawei Ascend (CANN), and AMD (ROCm).

Who it’s for

It is designed for researchers and developers working on Vision-Language-Action (VLA) models, robotic policy learning, and agentic AI who need a system capable of handling large-scale distributed training and real-world robot deployment.

Highlights

  • Broad Hardware Support: Compatible with NVIDIA, AMD, Huawei Ascend, and Moore Threads accelerators.
  • Extensive Simulator Integration: Supports Isaac Lab, LIBERO, RoboCasa, Metaworld, and others.
  • Real-World Capability: Includes RLinf-USER for online policy learning on physical robots like Franka and XSquare Turtle2.
  • System Optimizations: Features DynaRL for dynamic compute reallocation and FUSCO for accelerated MoE communication.
  • Versatile RL Support: Handles everything from offline IQL and SFT to online RL and GRPO fine-tuning for large models.

Related

  • Project
    alibaba/ROLL

    An efficient scaling library for reinforcement learning with LLMs that optimizes training and inference across large GPU clusters using a distributed Ray-based architecture.

  • Project
    areal-project/AReaL

    A large-scale asynchronous reinforcement learning system designed for training high-performance reasoning and agentic AI models.

  • Project
    nvidia-cosmos/cosmos-rl

    A scalable Reinforcement Learning framework specialized for Physical AI applications, providing a decoupled architecture for large-scale training workloads.

  • Project
    facebookresearch/ReAgent

    An end-to-end platform for applied reinforcement learning designed for large-scale distributed tasks and offline training without a simulator.

  • Project
    PrimeIntellect-ai/prime-rl

    A framework for large-scale asynchronous reinforcement learning designed to train 1T+ parameter MoE models across 1000+ GPUs using FSDP2 and vLLM.