Reflex: Real-Time Vision-Language-Action Control through Streaming Inference

Flow matching Vision-Language-Action (VLA) models promise precise continuous control, but their iterative denoising nature introduces fundamental incompatibilities with real-time robotics: global timestep injection invalidates KV-caching, forcing a choice between slow $O(N^2)$ re-computation or mathematically incorrect cache reuse. We present **Reflex**, a framework that enables *real-time streaming inference* for flow matching policies by exploiting the *Timestep-Invariance Property*---that perception encoders are functionally independent of the denoising loop. Reflex partitions the attention context into static, sliding, and dynamic regions, enabling $O(1)$ incremental cache updates while preserving full-batch-equivalent attention outputs for fixed inputs. To ensure stability under continuous high-frequency inference, we introduce *AdaRMSNorm*, an adaptive normalization layer that prevents BFloat16 numerical collapse by gating on flow phase. We further maximize throughput through an *async pipeline* that decouples visual encoding from action generation, combined with *operator fusion* that reduces kernel overhead. On LIBERO and Kinetix benchmarks, Reflex achieves a 2.58$\times$ inference speedup and 50Hz stable streaming, reducing reaction latency by up to 54\% and enabling efficient deployment without performance degradation.

Vision-language-action models allow robots to understand visual scenes and language instructions, but they are often too slow for real-time control. This is especially problematic for flow-matching policies, which generate actions through many iterative denoising steps. During this process, standard caching techniques used in language models cannot be directly reused because the model’s internal state changes with every denoising timestep. This paper introduces Reflex, a streaming inference system for real-time robotic control. The key idea is to separate parts of the model that remain stable, such as visual and language features, from parts that must change during action generation. Reflex reuses the stable computation safely, updates only the dynamic part, and runs perception and action generation asynchronously so the robot does not have to stop while planning. It also introduces an adaptive normalization layer to improve numerical stability during long-running mixed-precision inference. Experiments on simulation benchmarks and a real robot show that Reflex speeds up inference, reduces reaction latency, and maintains task performance, making flow-matching VLA models more practical for responsive robotic control.