Fast-Slow Asynchronous VLA Architecture Decouples 5 Hz Reasoning from 20 Hz Control in Closed-Loop Driving快慢异步VLA架构:5Hz推理与20Hz控制解耦实现闭环驾驶
S 1.7T11 sources1 个来源R7-research
Yun Li, Jiachen Gong and eight co-authors (arXiv:2607.15621v1, submitted 17 Jul 2026) propose an asynchronous fast-slow vision-language-action (VLA) architecture for end-to-end closed-loop driving, in which a frozen 7B vision-language backbone serves as a slow reasoning system decoupled from a separate fast control system.
The work targets a core mismatch: existing closed-loop agents invoke the LLM only on alternate simulation ticks and replay the previous command in between, meaning half of all control outputs ignore the newest observations.
The proposed system runs scene reasoning at 5 Hz and control actions at 20 Hz asynchronously, removing this replay compromise so LLM-based reasoning can inform driving decisions without stalling the vehicle's control loop.
Yun Li、Jiachen Gong 等10位作者（arXiv:2607.15621v1，2026年7月17日提交）提出一种用于端到端闭环驾驶的快慢异步视觉-语言-动作（VLA）架构，其中冻结的7B视觉语言主干网络作为慢速推理系统，与独立的快速控制系统解耦。