REA runtime coordinates slow planners with fast robot policies
Original title: REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
On 3,000 LIBERO-Pro episodes it reports 95.9% success, 25% fewer control steps, and 17% less control time than waiting on the brain.
The paper introduces REA, a routed embodied-agent runtime that keeps a deliberative foundation-model planner, a router, and fast reactive policies from blocking one another. A route memory, a boundary-triggered judge, and a shared-budget escalation ladder decide when to consult the slow model. On 3,000 episodes across 60 LIBERO-Pro tasks, REA reports 95.9% success versus 94.7% for the best single policy in its pool. A separate 100-episode RoboTwin 2.0 study reports 84% with VLA policies and 86% with VLA plus code, against 80% for MEDex, with additional transfer to BEHAVIOR-1K and a Unitree G1.
- The golden rule is that no running policy waits on the foundation-model brain.
- LIBERO-Pro covered 3,000 test episodes and 60 tasks.
- RoboTwin results are from a separate 100-episode study.
- REA also ran on BEHAVIOR-1K and a real Unitree G1 humanoid.
- Learned VLA policies and programmatic code policies share the same runtime.
- Vendors
- Unitree
- Products
- REAUnitree G1
- AI models
- MEDex
From the original article
1 min read at huggingface.co
Foundation models are becoming remarkably reliable at high-level reasoning, yet they remain too slow and costly to serve as fine-grained controllers. Conversely, fast policies like VLAs and whole-body controllers offer more efficient reflexes, but struggle to generalize.
By Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.