LIRL: Latent Imagination-Based Reinforcement Learning for Efficient Coverage Path Planning

Coverage Path Planning (CPP) in unknown environments presents unique challenges that often require the system to maintain a symmetry between exploration and exploitation in order to efficiently cover unknown areas. This paper introduces latent imagination-based reinforcement learning (LIRL), a novel framework that addresses these challenges by integrating three key components: memory-augmented experience replay (MAER), a latent imagination module (LIM), and multi-step prediction learning (MSPL) within a soft actor–critic architecture. MAER enhances sample efficiency by prioritizing experience retrieval, LIM facilitates long-term planning via simulated trajectories, and MSPL optimizes the trade-off between immediate rewards and future outcomes through adaptive n-step learning. MAER, LIM, and MSPL work within a soft actor–critic architecture, and LIRL creates a dynamic equilibrium that enables efficient, adaptive decision-making. We evaluate LIRL across diverse simulated environments, demonstrating substantial improvements over state-of-the-art methods. Through this method, the agent optimally balances short-term actions with long-term planning, maintaining symmetrical responses to varying environmental changes. The results highlight LIRL’s potential for advancing autonomous CPP in real-world applications such as search and rescue, agricultural robotics, and warehouse automation. Our work contributes to the broader fields of robotics and reinforcement learning, offering insights into integrating memory, imagination, and adaptive learning for complex sequential decision-making tasks.