Robot news usually favors new brains, world models and polished demos. RLinf v0.3 is less theatrical, but it sits in the part of embodied AI that often decides whether a demo can become a repeatable workflow.
Released on July 16 by Infinigence AI, Tsinghua University and collaborators, the update connects data collection, supervised fine-tuning, reinforcement learning, evaluation and real-robot deployment.
“RLinf is a flexible and scalable open-source RL infrastructure designed for Embodied and Agentic AI.”
A pipeline, not a single feature
The v0.3 release adds support for six embodied models, five simulators or environments, three teleoperation methods, three real-world platforms and two end-effectors. Those numbers matter because embodied AI teams often lose time rebuilding glue code for every robot, camera, gripper, simulator and dataset format.
The project is trying to make the training path portable. A researcher can collect demonstrations with a SpaceMouse, VR or GELLO, run SFT, continue with PPO, GRPO, DSRL, RECAP or SAC-Flow, then deploy to platforms such as Franka, GimArm or DOS-W1.
The system layer is the harder story
RLinf v0.3 also adds reward and value model components, SGLang serving, decoupled environment execution, torch.compile acceleration, rollout-training overlap, upgraded weight synchronization and FSDP full offload. The arXiv system paper reports 1.07x to 2.43x end-to-end training throughput gains across reasoning RL and embodied RL tasks.
The release also adds end-to-end Ascend CANN / torch-npu support and continues work on AMD ROCm and Musa. For Chinese robotics teams, that matters because RL workloads are sensitive to scheduling, memory pressure and environment execution, not only raw accelerator speed.
The test is independent reproduction
GitHub shows more than 4,100 stars, 600 forks and 100 contributors. QbitAI also notes that Isaac Lab has accepted RLinf as a training engine for embodied models, and that a medical-device assembly task with NVIDIA appeared at GTC 2026.
Those are useful signals, but they are not deployment proof. The next checks are whether outside teams can reproduce v0.3 examples, whether non-NVIDIA hardware keeps pace with the main branch, and whether success-rate gains survive real robot differences.
If those checks hold, RLinf v0.3 will be more than a version bump. It will move embodied AI training one step away from isolated demos and toward portable engineering.
Sources: QbitAI, RLinf GitHub repository, RLinf v0.3 Release Notes, arXiv:2509.15965, CocoLoop; checked the v0.3 release node, six models, five simulation environments, three teleoperation methods, three real-robot platforms, two end-effectors, 4,100-plus stars, 600-plus forks, 100-plus contributors and the 1.07x-2.43x training-throughput figure.