Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation

2026-08-14

AI Summary

This study addresses the lack of evaluation and inefficient reasoning in existing Vision-Language-Action (VLA) models for dynamic, reaction-critical scenarios by proposing ReflexBench and ReflexVLA. The proposed architecture achieves high efficiency without large-scale pretraining through latent future prediction, multi-frame temporal fusion, and CUDA Graph optimization, while supporting decoupled simulation and configurable latency evaluation. Experimental results demonstrate that ReflexVLA significantly enhances performance on dynamic tasks while maintaining accuracy in static settings. Furthermore, real-world deployment validates its effectiveness under low-latency conditions, establishing a novel paradigm for reactive manipulation.

Abstract

Vision-Language-Action (VLA) models have recently achieved promising performance in robotic manipulation. However, existing benchmarks mainly evaluate generalization on static manipulation tasks and largely overlook dynamic interaction scenarios. To address this gap, we present ReflexBench, a benchmark for reaction-critical manipulation. ReflexBench contains six dynamic tasks and introduces an evaluation framework that decouples simulator stepping from robot control while supporting configurable latency under synchronous and asynchronous inference. Building upon ReflexBench, we propose ReflexVLA, an efficient VLA model designed for reaction-critical manipulation without large-scale robot-data pretraining. ReflexVLA enhances temporal reasoning through latent future prediction and multi-frame temporal fusion within the vision backbone, while reducing deployment latency through batched visual encoding and CUDA Graph replay. Experiments show that ReflexVLA consistently improves dynamic manipulation performance while maintaining competitive accuracy on standard static manipulation benchmarks, and real-world experiments further demonstrate its effectiveness under practical deployment conditions. Project website: https://reflexvla.github.io

Problem

• Vision-Language-Action Models
• Reaction-Critical Manipulation
• Dynamic Interaction
• Benchmark
• Deployment Latency

Innovation

• Vision-Language-Action Models
• Reaction-Critical Manipulation
• Latent Future Prediction
• Multi-frame Temporal Fusion
• CUDA Graph Replay