REA: A Routing-Based Embodied Agent Runtime That Asynchronously Coordinates a Slow 'Brain' and a Fast 'Cerebellum'

Problem

Foundation models are good at planning and judgment but are too slow and expensive to use as low-level controllers. Fast policies such as VLAs generalize poorly. Combining the two is also hard because the two approaches were built on different simulators and evaluation criteria, making it difficult to compare them under the same conditions. In addition, setups where an agent calls a VLA bring the robot to a stop at every decision, and setups where a VLA calls an agent on errors get called repeatedly for the same error and miss failures that occur without an error.

Method

The system is split into System 2 (brain: policy generation, advice and completion judgment), System 1 (router: decides which policy to run and when, without producing actions directly) and System 0 (code policies, VLAs and whole-body controllers). There is one rule: 'a running policy does not wait for the brain.' Route memory stores the per-policy success rates and step counts measured on development episodes, and at test time a policy is chosen by looking up the table. The judge gate asks a VLM whether the task is done each time a policy finishes. If it is done, the system stops to preserve the completed result; otherwise it moves to the next stage. The escalation ladder runs the next policy from the current state within a single step and time budget.

Results

• On 60 LIBERO-Pro tasks and 3,000 episodes, REA-M reaches a 95.9% success rate, higher than the best single baseline, Agents SDK + GPT-6 (94.7%, p=0.004), and 1 percentage point from the post-hoc oracle (96.9%).
• Steps per success are 1,003, 25% fewer than Agents SDK (1,343), and control time is 17% shorter. It is also faster than configurations that wait for LLM advice.
• In a 100-episode experiment on RoboTwin 2.0, it reaches 84% with a VLA policy pool and 86% with a VLA+code pool, higher than the best single-VLA baseline, MEDex (80%).
• The judge gate raised the success rate by 3.1 percentage points over an exception-based gate (p=0.022), but took 2.1 times as long per success. Show-Harness, which asks a VLM about every action, reached only a 2.6% success rate.
• On BEHAVIOR-1K it scored 27/50, higher than Agents SDK (25/50), but the difference was not statistically significant. Experiments on a real Unitree G1 succeeded in 8 of 10 trials.

Limitations

The LIBERO-Pro evaluation includes only the task and swap variants and excludes the Long suite and the object, language and environment variants. Building Route memory requires running development episodes for each task in advance. The difference on BEHAVIOR-1K was not significant, and the real-robot experiment is small at 10 trials. Using the judge gate increases time cost.

Why it matters

It makes it possible to compare LLM agents, VLAs and code policies under the same execution contract and the same evaluation conditions. It shows a practical division of labor in which heavy models are used only for offline policy preparation and judgment when needed, while real-time control is left to fast policies.

Key terms

VLA (Vision-Language-Action) model

A model that takes images and language instructions as input and directly outputs robot actions. It is fast, but its performance tends to drop outside its training distribution.

Route memory

A lookup table that stores per-policy success rates and costs measured on development episodes, used at test time to pick policies without calling an LLM.

Silent failure

A case where a policy finishes normally without raising an error but has not actually completed the task. Exception-based recovery cannot catch it.

Code-as-policy

An approach in which a coding agent writes a control program that calls perception, grasping and inverse kinematics tools, and that program is used as the robot policy.

Original

• arXiv abstract
• PDF
• HTML full text
• Project
• Hugging Face

Authors: Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang, Haoyuan Chen, Zhenglin Wan, Hexin Wang, Wangbo Zhao, Yang You · Posted 2026-10-30 · License CC BY-NC-SA 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.