Wuxian Qionxin Releases Open-Source Embodied Intelligence Framework RLinf v0.4

On October 28, 2026, AI-native infrastructure company Wuxian Qionxin, together with Tsinghua University and other institutions, released RLinf v0.4, an open-source framework for large-scale reinforcement learning training in embodied intelligence. The project is described as the first open-source framework of its kind targeting embodied intelligence at scale.

The new version focuses on reinforcement learning in real physical scenarios, with upgrades in four areas: modular robot systems, efficient real-robot algorithms, full-modality model support, and domestic computing power. Hardware such as robotic arms, dexterous hands, cameras, and mobile bases is abstracted into plug-and-play modules, so upper-layer task and policy code remains unchanged when the robot body or configuration is replaced. New algorithms include online policy distillation and asynchronous real-time control. Online policy distillation improves model task success from 67% to over 96% in 30 training steps, while asynchronous real-time control reduces inference waiting time to about 2 milliseconds.

RLinf v0.4 also integrates five frontier world action models and VLA models, including NVIDIA Cosmos 3. For computing power adaptation, the framework adds support for Kunlunxin, Biren Technology, Moore Threads, and Intel XPU. It now runs seamlessly on NVIDIA, AMD, and six major domestic and heterogeneous accelerators. On Huawei Ascend, the end-to-end computing efficiency of ten mainstream embodied models exceeds 75% of A100 performance.

On the ecosystem side, RLinf has been included in the official PyTorch ecosystem and is the officially integrated reinforcement learning training engine in NVIDIA Isaac Lab v3.0. The team had two research papers accepted at CoRL 2026 and one paper accepted at NSDI 2027.