An exploration-driven framework for path planning in complex buildings using improved MADDPG

Abstract

Multi-agent deep reinforcement learning (MADRL) methods have been extensively applied to crowd evacuation in complex building environments. However, the intricate architecture of modern buildings and high population densities hinder agent exploration efficiency, while existing approaches struggle to overcome the challenges posed by sparse rewards. To tackle these issues, this study proposes the Intrinsic Curiosity Distillation Multi-Agent Deep Deterministic Policy Gradient (ICD-MADDPG) algorithm, an exploration-driven framework for path planning in complex buildings. First, the ICD-MADDPG algorithm introduces a curiosity mechanism by integrating the Intrinsic Curiosity Module (ICM) and Random Network Distillation (RND), thereby refining the reward mechanism and significantly enhancing exploration efficiency. Next, the feature extraction process is enhanced using a Long Short-Term Memory (LSTM) network, enabling the model to effectively capture temporal dependencies in dynamic environments. Finally, a two-layer evacuation mechanism is adopted, where the crowd is divided into groups consisting of leaders and followers. Leaders utilize the ICD-MADDPG algorithm for global evacuation path planning, while followers employ the Reciprocal Velocity Obstacle (RVO) algorithm to follow the leaders and avoid collisions efficiently. Experimental results demonstrate that the ICD-MADDPG algorithm achieves superior rewards, improves evacuation efficiency, and effectively mitigates congestion. This framework provides a robust theoretical basis for optimizing evacuation strategies and offers practical value for intelligent building systems and emergency response planning.