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(N2) re-computation or mathematically incorrect cache reuse. We present Reflex, a framework that enables textitreal-time streaming inference for flow matching policies by exploiting the textitTimestep-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 textitAdaRMSNorm, an adaptive normalization layer that prevents BFloat16 numerical collapse by gating on flow phase. We further maximize throughput through an textitasync pipeline that decouples visual encoding from action generation, combined with textitoperator fusion that reduces kernel overhead. On LIBERO and Kinetix benchmarks, Reflex achieves a 2.58× inference speedup and 50Hz stable streaming, reducing reaction latency by up to 54% and enabling efficient deployment without performance degradation.