REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
What it is
REA is a proposed runtime for embodied agents that coordinates three layers: a deliberative System 2 (foundation models), an embodied router System 1, and reactive System 0 policies, under the rule that no running policy waits on the brain. It uses a route memory storing each policy's measured success, a boundary-triggered judge gate that consults the brain during policy execution, and a shared-budget escalation ladder for bounded recovery. The paper reports 95.9% success on 3,000 test episodes across 60 LIBERO-Pro tasks (vs. 94.7% for the best single-agent baseline in its pool) with 25% fewer control steps and 17% less control time, plus 84-86% success on a 100-episode RoboTwin 2.0 study and transfer to BEHAVIOR-1K and a real Unitree G1 humanoid.
The problem it solves
Pairing a deliberative foundation model with fast reactive policies is hard: an agent that calls the VLA in the agentic loop stalls fast execution, while a VLA that calls the agent on errors re-triggers fallbacks repeatedly or misses silent failures, and existing methods are built and scored on incompatible stacks.
What you could use it for
- Coordinating a slow foundation-model planner with fast VLA or whole-body controllers in robot manipulation tasks
- Bounded error recovery in embodied agents without repeatedly re-triggering fallbacks or missing silent failures
- Benchmarking mixed learned and programmatic policies within a single runtime on LIBERO-Pro or RoboTwin 2.0
- Deploying asynchronous brain-cerebellum control on humanoid hardware such as the Unitree G1
Why it matters
If the reported results hold, it would let embodied systems combine general-purpose foundation-model reasoning with fast reactive policies without the latency and cost penalties of querying the brain on every failure, potentially making hybrid brain-cerebellum control practical on real robots.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.