Reflex: Real-Time VLA 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 $\mathbf{\text{Reflex}}$, a framework that enables $\mathit{\text{real-time}}\mathit{\text{streaming}}\mathit{\text{inference}}$ for flow matching policies by exploiting the $\mathit{\text{Timestep-Invariance}}\mathit{\text{Property}}$ – 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 $\mathit{\text{AdaRMSNorm}}$, an adaptive normalization layer that prevents BFloat16 numerical collapse by gating on flow phase. We further maximize throughput through an $\mathit{\text{async}}\mathit{\text{pipeline}}$ that decouples visual encoding from action generation, combined with $\mathit{\text{operator}}\mathit{\text{fusion}}$ 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.