Improving the Efficiency of Collaboration Between Humans and Embodied AI Agents in 3D Virtual Environments

This study proposes a human-in-the-loop dynamic graph-based planning framework designed to elevate LLM-based Embodied Agents from simple tools to trustworthy collaborative partners. To achieve this, we address the trade-off between the structural rigidity of plan-centric approaches and the instability of reactive methods. The framework utilizes a Directed Acyclic Graph (DAG) with AND/OR nodes to ensure robustness while maintaining flexibility. Critically, the agent features an Automated Recovery Mechanism for self-correction and a Dynamic Modification Mechanism that employs Relevance Analysis to effectively translate human interventions (Switch, Add, Delete) into graph updates. Comparative experiments in Minecraft with 30 participants validated the method’s effectiveness. The proposed agent (Agent B) outperformed the reactive baseline (Agent A), reducing mission completion time by 9.3%. Notably, the agent demonstrated high instruction compliance and reduced user frustration by approximately 20%, leading to statistically higher satisfaction scores (PSSUQ). These results confirm that by ensuring planning robustness and responsiveness, the proposed framework successfully enables agents to function as trustworthy partners in complex environments.