Multi-Expert Workflow Robot Planning Model Based on Routed Cyclic Optimization

Jun Wang

Keywords: Multi-Expert Workflow, Routed Cyclic Optimization, Embodied AI Planning, Planner-Advisor Collaboration

Abstract:

This paper proposes a "Multi-Expert Workflow based on Routed Cyclic Optimization" model for the EAI Challenge. Our core innovation lies in designing a dynamic dual-routing architecture:
First, a "Classifier Expert" analyzes the input task and precisely categorizes it into one of the four main challenge tasks (Goal Interpretation, Subgoal Decomposition, Action Sequencing, and State Transition Modeling).
Subsequently, the problem is routed to a specialized "Planner-Advisor" expert group that is highly optimized for this specific task.
The workflow employs multi-round iterative cycles for refinement, incorporating dynamic cycle control and a "Final Review" mechanism for persistent errors.
Our method demonstrates strong generalization capabilities and performance improvements in both the BEHAVIOR and VirtualHome environments, achieving an overall score of 70.38 across all tasks.
These results validate the effectiveness of our "classification-routing-optimization" workflow in addressing complex embodied AI planning problems.

Submission Number: 12