Contextual Knowledge Sharing in Multi-Agent Long-Horizon Planning Settings with Centralized Communication and Coordination

Abstract

Embodied multi-agent systems, comprising autonomous agents interacting within shared environments, enable intelligent, collaborative solutions for tasks requiring real-time coordination and adaptability. While applications span diverse fields, from disaster response to healthcare, planning in these systems remains challenging due to partial egocentric observations and limited environmental awareness. This work addresses these challenges by introducing a software module that synthesizes a shared world state from individual agent views, maintaining spatial information about objects and agents to support more effective joint action planning. Integrated into the LLAMAR framework, this module aims to improve planning accuracy and efficiency. The proposed approach is evaluated using metrics such as success rate, transport efficiency, and coverage performance. Our evaluation demonstrates that utilizing a perfect (oracle-generated) world state significantly enhances planning effectiveness. Notably, under these ideal conditions, the success rate of the LLAMAR planner improved by over 16%. These findings underscore the critical impact of accurate world state representation on multi-agent performance and highlight the potential for significant advancements in collaborative task execution in dynamic, unstructured settings.