RLinf is an open-source reinforcement learning infrastructure framework for embodied and agentic AI, designed to train vision-language-action (VLA) models and LLM agents with RL at scale. It has earned about 5,500 stars on GitHub and is actively developed, with its latest updates in 2026.
Training embodied agents with reinforcement learning is an infrastructure problem as much as an algorithms problem: simulation rollouts, policy inference and gradient updates compete for the same GPUs, and naive pipelines leave hardware idle. RLinf addresses this with a flexible distributed architecture that can colocate or disaggregate rollout, inference and training, scheduling them across clusters to keep utilization high. It supports RL fine-tuning of VLA models for robot manipulation, integrates with popular simulators and benchmarks, and also covers agentic AI workloads such as tool-using LLM agents. Multiple RL algorithms are supported, and the framework emphasizes reproducible configurations and scalable throughput, so research teams can go from small experiments to large multi-node runs without rewriting their training stack. A key design goal is flexibility in execution layout: researchers can share GPUs between generation and training or fully separate them, and choose the arrangement that fits their model sizes and simulator costs. The team publishes documentation, examples and benchmark configurations to make large embodied-RL runs reproducible.
RLinf is completely free and open source under the Apache-2.0 license. You supply your own compute and simulators.
RLinf is for robotics and AI research teams training embodied agents or RL-fine-tuning VLA and LLM models who need training infrastructure that scales without wasting GPU time.