REA Links Foundation-Model Reasoning With Fast Robot Policies
Pengfei Zhou and seven coauthors submitted an arXiv paper introducing REA, a routed embodied-agent runtime that coordinates a deliberative foundation-model layer, an embodied router and fast reactive policies. Its design keeps running policies from waiting on the model and uses route memory, a boundary-triggered judge gate and a shared-budget escalation ladder. The paper reports 95.9% success over 3,000 LIBERO-Pro episodes spanning 60 tasks, 25% fewer control steps and 17% less control time than the best single-agent baseline in its policy pool. A separate 100-episode RoboTwin 2.0 study reported 84% success with VLA policies and 86% with VLA plus code, versus 80% for the MEDex baseline. The paper also reports results on BEHAVIOR-1K and a real Unitree G1 humanoid.
The paper presents REA as a concrete runtime for combining high-level foundation-model reasoning with fast robot control, addressing the difficulty of coordinating these components asynchronously so that model calls do not stall policy execution.
Computer Science > Robotics
arXiv:2610.42053 (cs)
[Submitted on 30 Oct 2026]
Title:REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
Authors:Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang, Haoyuan Chen, Zhenglin Wan, Hexin Wang, Wangbo Zhao, Yang You
View a PDF of the paper titled REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control, by Pengfei Zhou and 7 other authors