REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control

REA is a routed embodied agent runtime that coordinates a slow System 2 brain, a router and fast System 0 policies so that no running policy waits on the brain.

Foundation-model brains are reliable at high-level reasoning but slow and costly as fine-grained controllers, while fast VLA and whole-body policies generalize poorly. Pairing a deliberative brain with a reactive cerebellum is hard because methods are built and scored on incompatible stacks and asynchronous coordination is difficult. An agent that calls the fast policy stalls execution, while a fast policy that calls the agent on errors re-triggers fallbacks or misses silent failures.

REA is a unified runtime with a System 2 brain, a System 1 embodied router and System 0 fast policies under a golden rule that no running policy waits on the brain. The brain generates policies offline or online and provides advice or judgments. Route memory stores each policy's measured per-task success and cost from development episodes so dispatch does not require System 2 reasoning. A boundary-triggered judge gate consults System 2 concurrently to detect completion and silent failures, stopping when done or escalating when not. A shared-budget escalation ladder dispatches the next policy from the current state under one step and time budget. The same interfaces and aligned evaluation protocol cover code policies, VLA policies and whole-body controllers across simulators and real robots.

• On the full LIBERO-Pro set, REA-M reaches higher success than the best single-agent baseline in its own policy pool and uses fewer control steps per success. §4.2
• Route memory reduces control time versus the blocking advisor and versus Agents SDK, while matching or improving success. §4.3
• On ASPIRE Fix-loop policies, REA-M lifts success and the judge gate adds further success at higher wall time per success. §4.4
• In the RoboTwin 2.0 VLA and code study, REA improves over the strongest single VLA baseline MEDex and adding code raises randomized successes. §4.6
• On BEHAVIOR-1K, REA solves the most test episodes among the compared methods. §4.7
• On the real Unitree G1 walk-and-push-chair task, REA has the highest operator-judged success. §4.8

• The RoboTwin VLA and code study notes that historical baselines and the VLA extension were run in different windows, and MEDex uses a separate host, so those comparisons carry a window and host caveat (§4.6).
• The judge-gate attribution is approximate because REA-M's own run stands in for REA-V's first run, so the split into rescued failures, undone successes and replay variance is only approximate (§4.4, §B.6).
• The BEHAVIOR-1K margin over the Agents SDK agent is not significant and comes from one task (§4.7).
• The AlohaMini2 real-robot studies use ten valid trials per method and retain different days or session settings; REA also has longer mean recorded time in both studies, and early termination of failed trials affects runtime averages (§4.9, §D.3).

• REA separates a System 2 brain, a System 1 embodied router and System 0 fast policies under one runtime. §1
• Route memory chooses the first policy and ladder from development success and cost, without model tokens at dispatch. §4.2
• A boundary-triggered judge gate consults System 2 concurrently to decide whether the task is finished, escalating if not and stopping if done. §1
• The escalation ladder runs under one shared step and wall budget, and escalations do not extend the episode. §B.1
• The aligned evaluation protocol uses the same task-seed pairs, initial-state semantics, horizon and wall budget within matched comparisons. §3.3
• Route memory used 10 development seeds per program and no model tokens, while generation pipelines used 1,249–1,328 development episodes and 29–89M input tokens. §4.2
• Direct brain-as-controller policies solved 1 of 1,987 one-token episodes and 0 of 600 Cartesian IK episodes. §1
• On AlohaMini2, REA completes 10/10 trials versus ASPIRE 8/10, Codex 5/10 and Agents SDK 4/10; a separate study routes between pi0.5 and Code for 8/10 versus 3/10 for pi0.5 alone. §4.9