EVENT DOSSIER
On September 7, 2026, arXiv released RL-VLA$^3$. This is a fully asynchronous distributed reinforcement learning framework designed specifically for visual-language-action (VLA) model training. Through a dynamic batch scheduling mechanism and flexible environment partitioning strategies, this framework achieves fine-grained asynchronous interaction between simulation, reasoning, and training components, aiming to address the mismatch issues caused by traditional synchronous designs in high-latency physical simulators. Experimental results show that RL-VLA$^3$ achieves up to 85.2% higher throughput compared to synchronous baselines, while maintaining consistent sample efficiency. Its scalability has been verified on GPUs ranging from 8 to 256. It is claimed to be the first fully asynchronous reinforcement learning training framework tailored for VLA training system-level challenges.
arXiv:2602.05765v3 发布 RL-VLA$^3$，这是一个专为视觉 - 语言 - 动作（VLA）模型训练设计的完全异步分布式强化学习框架。该框架通过动态批处理调度器和灵活的环境分片策略，实现了仿真、推理与训练组件之间的细粒度异步交互，以解决传统同步设计在物理模拟器高延迟场景下的不匹配问题。实验表明，RL-VLA$^3$ 相比同步基线吞吐量提升最高达 85.2%，且保持样本效率一致，并已在 8 至 256 张 GPU 上验证了可扩展性。据称，这是首个针对 VLA 训练系统级挑战定制的完全异步强化学习训练框架。