Hugging Face Releases REA: Asynchronous Brain-Cerebellum Control Framework for Efficient Embodied Agent Collaboration

Summary:Hugging Face has introduced REA (Routed Embodied Agent), a novel runtime framework for embodied agents that addresses the challenge of asynchronous coordination between a deliberative brain (foundation models) and a reactive cerebellum (fast policies). REA employs a route memory, a boundary-triggered judge gate, and a shared-budget escalation ladder to enable policy execution without waiting for the brain, significantly improving control efficiency. Evaluations on LIBERO-Pro and RoboTwin 2.0 tas

---

1. Introduction

In the field of artificial intelligence, foundation models excel at high-level reasoning tasks but are computationally expensive and slow, making them unsuitable for real-time control. Conversely, fast policies like VLA and whole-body controllers are efficient but lack generalization capabilities. Combining foundation models with fast policies to create a system similar to the human brain and cerebellum is a natural solution. However, existing methods face two main challenges:

1. Incompatible Technical Stacks: Different methods are built on incompatible technical stacks, making it difficult to work together.
2. Difficult Asynchronous Coordination: Coordinating between the agent and fast policies in an asynchronous environment often leads to delays or errors.

2. Core Mechanisms of REA

REA (Routed Embodied Agent) addresses these issues through the following mechanisms:

• Route Memory: Pre-stores the success rate of each policy to avoid querying the foundation model every time a failure occurs.
• Boundary-Triggered Judge Gate: Triggers a query to the foundation model based on boundary conditions during policy execution to ensure decision accuracy.
• Shared-Budget Escalation Ladder: Provides a limited recovery budget for policy execution to prevent infinite loops or repeated triggering.

3. Experimental Results and Analysis

REA achieved a 95.9% success rate in 3,000 test scenarios of LIBERO-Pro, surpassing the best single-agent baseline (94.7%) while reducing control steps by 25% and execution time by 17%. In the 100 test scenarios of RoboTwin 2.0, REA achieved an 84% success rate with VLA policies and 86% with VLA+code, compared to 80% for the strongest single VLA baseline MEDex. These results demonstrate that REA outperforms existing methods in complex tasks.

4. Practical Applications and Deployment Recommendations

The REA framework not only performs well in simulated environments but has also been successfully deployed on BEHAVIOR-1K and the real Unitree G1 humanoid robot, showcasing its potential for practical applications. For developers, the following characteristics make REA an ideal choice for building efficient agents:

• Modular Design: REA's components can be independently optimized and extended, facilitating integration into existing systems.
• Cross-Task Adaptability: REA can adapt to different tasks and environments, supporting multiple policy types.
• Efficient Resource Utilization: By reducing reliance on foundation models, REA lowers computational costs and improves real-time performance.

5. Conclusion

The REA framework, with its innovative asynchronous coordination mechanism, enables efficient collaboration between foundation models and fast policies, significantly improving the control efficiency and adaptability of agents. Its excellent performance in multiple benchmarks proves its powerful capabilities in complex tasks, paving the way for new directions in AI agent development.

---

Source: Hugging Face Daily Papers (2026-10-30)

— END —

Tags: #Hugging Face Follow #Intelligent Agents Follow #Asynchronous Control Follow #REA Follow