RLinf

Why we included this project

RLinf is worth a look if you are training robot policies or open-ended agents with reinforcement learning and would rather not assemble the pipeline yourself. It covers the training lifecycle from demonstration data collection and supervised fine-tuning through RL to deployment, and it handles a range of embodied setups, from LIBERO manipulation benchmarks and offline D4RL tasks to real robot arms and dexterous hands. Support for vision-language-action models such as GR00T and the PyTorch reimplementations of the π₀ / OpenVLA line is built in, and it plugs into simulators like Genesis and Polaris, with system-level optimizations that noticeably cut rollout latency. Rather than prescribing a single algorithm, it lets you wire up data collection, policy updates, and simulator backends on your own terms. That makes it a practical base for researchers and small robotics teams looking to reproduce or extend recent embodied RL work.