RLinf
An open-source reinforcement-learning infrastructure for embodied and agentic AI, used to train robotics and vision-language-action models at scale with GRPO and other RL algorithms.
RLinf is a flexible, scalable reinforcement-learning infrastructure designed for embodied and agentic AI training. Its name stands both for infrastructure, describing its role as a backbone for large-scale training, and for infinite, describing its aim of supporting open-ended learning and continuous generalization.
It has been adopted as the reinforcement-learning training backbone for Isaac Lab and was accepted into the PyTorch Ecosystem, and it supports hardware including NVIDIA, Moore Threads, Huawei Ascend and AMD accelerators. Supported workloads include GRPO training for large mixture-of-experts models, supervised fine-tuning and evaluation for vision-language-action models, and reinforcement learning for video-generation models.