REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
REA keeps robot policies running without waiting for the reasoning model, reporting higher success with less control time.
The runtime combines a deliberative model, a router and fast reactive policies. It uses stored policy success measurements, a boundary-triggered judge gate and a shared recovery budget to coordinate them asynchronously. Across 3,000 episodes on 60 LIBERO-Pro tasks, the authors report 95.9% success versus 94.7% for the strongest single-agent baseline in its policy pool, with 25% fewer control steps and 17% less control time. In a separate 100-episode RoboTwin 2.0 study, success reached 84% with VLA policies and 86% with VLA plus code, compared with 80% for the strongest single VLA baseline.
HF Daily Papers' note
The runtime combines a deliberative model, a router and fast reactive policies. It uses stored policy success measurements, a boundary-triggered judge gate and a shared recovery budget to coordinate them asynchronously. Across 3,000 episodes on 60 LIBERO-Pro tasks, the authors report 95.9% success versus 94.7% for the strongest single-agent baseline in its policy pool, with 25% fewer control steps and 17% less control time. In a separate 100-episode RoboTwin 2.0 study, success reached 84% with VLA policies and 86% with VLA plus code, compared with 80% for the strongest single VLA baseline.
HF Daily Papers' note
score 4