文章背景与核心概要
视觉-语言-动作（VLA）模型近年来在机器人操作领域取得了显著进展，但现有的基准测试几乎全部集中于静态任务评估，严重忽视了需要快速响应的动态交互场景。为了填补这一空白，本文推出了 ReflexBench——一个专为反应关键型操作设计的全新基准测试，包含六个动态任务，并引入了解耦仿真器步进与机器人控制的评估框架，支持同步和异步推理下的可配置延迟。
基于 ReflexBench，作者进一步提出了 ReflexVLA，这是一种专为反应关键型操作设计的高效 VLA 模型，无需进行大规模机器人数据预训练。ReflexVLA 通过在视觉骨干网络中引入潜在未来预测（Latent future prediction）和多帧时间融合（Multi-frame temporal fusion）来增强时间推理能力；同时，通过批处理视觉编码（Batched visual encoding）和 CUDA Graph 重放（CUDA Graph replay）大幅降低部署延迟。实验表明，ReflexVLA 在动态操作任务上超越了现有模型，同时在标准静态基准测试中保持了强劲的竞争力，真实世界的实验也进一步验证了其在实际部署条件下的有效性。
Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation
Authors: Yuxuan Chen, Wanruo Zhang, Xiao Li
Subjects: Robotics (cs.RO), Artificial Intelligence (cs.AI)
arXiv: 2608.14379 [cs.RO]
Submitted: August 14, 2026
Project Website: https://reflexvla.github.io
📌 Summary
Vision-Language-Action (VLA) models have revolutionized robotic manipulation, but current benchmarks evaluate them almost exclusively on static tasks, ignoring dynamic interaction scenarios. To bridge this gap, this paper introduces ReflexBench, a novel benchmark containing six dynamic tasks with an evaluation framework that decouples simulator stepping from robot control under both synchronous and asynchronous inference.
Building upon ReflexBench, the authors propose ReflexVLA, an efficient VLA model designed specifically for reaction-critical manipulation without requiring large-scale robot-data pretraining. ReflexVLA introduces: * Latent future prediction and multi-frame temporal fusion within the vision backbone to enhance temporal reasoning. * Batched visual encoding and CUDA Graph replay to drastically reduce deployment latency.
Experiments show that ReflexVLA outperforms existing models on dynamic manipulation while remaining competitive on standard static benchmarks, with real-world validation confirming its practical deployment capabilities.
📑 Abstract
视觉-语言-动作（VLA）模型近期在机器人操作领域取得了令人瞩目的性能。然而，现有的基准测试主要评估静态操作任务上的泛化能力，很大程度上忽视了动态交互场景。为了解决这一差距，我们提出了 ReflexBench，这是一个针对反应关键型操作的基准测试。ReflexBench 包含六个动态任务，并引入了一个评估框架，该框架将仿真器步进与机器人控制解耦，同时支持在同步和异步推理下进行可配置的延迟。基于 ReflexBench，我们提出了 ReflexVLA，这是一种专为反应关键型操作设计的高效 VLA 模型，无需进行大规模机器人数据预训练。ReflexVLA 通过视觉骨干网络中的潜在未来预测和多帧时间融合来增强时间推理能力，同时通过批处理视觉编码和 CUDA Graph 重放来减少部署延迟。实验表明，ReflexVLA 在持续提升动态操作性能的同时，在标准静态操作基准测试中保持了具有竞争力的准确率，真实世界实验进一步证明了其在实际部署条件下的有效性。
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.
🔗 Quick Links
快速链接