RLinf

RLinf (Reinforcement Learning Infrastructure) is an open-source, flexible and scalable reinforcement learning (RL) infrastructure designed for embodied and agentic AI. The 'inf' stands for both 'Infrastructure' and 'Infinite', emphasizing support for open-ended, continuous learning. It supports diverse RL training workflows (PPO, GRPO, SAC) while hiding the complexity of distributed programming, letting users scale training across many GPU nodes without code changes. A key innovation described in its paper is macro-to-micro flow transformation (M2Flow) together with a hybrid execution mode, which achieves up to 2.43x higher throughput than existing frameworks for embodied RL. RLinf integrates multiple backends (FSDP + HuggingFace/SGLang/vLLM and Megatron + SGLang/vLLM) and supports training of VLA models, VLMs and world models in simulation and the real world. The framework originated in 2025 as a collaboration among researchers from Tsinghua University, Infinigence-AI, Peking University and Beihang University, and is released under the Apache 2.0 license. RLinf has been welcomed into the PyTorch Ecosystem (Q3 2026) and adopted as the RL training infrastructure in Isaac Lab v3.0.0.