Autonomous navigation with minimal sensors in dynamic warehouse environments: a multi-agent reinforcement learning approach with curriculum learning enhancement

Abstract This paper introduces an innovative approach to augmenting the capabilities of autonomous mobile robots (AMRs) in dynamic, multi-agent warehouse environments. We propose an integrated system that combines Multi-Agent Deep Deterministic Policy Gradient (MADDPG) a cutting-edge multi-agent reinforcement learning (MARL) algorithm with Curriculum Learning (CL). This hybrid method is specifically tailored to meet the complex demands of warehouse operations. Our approach is distinguished by the development of decentralized policies within a centralized training framework, progressively incorporating complexity through a carefully structured curriculum. This strategy enables autonomous agents to master sophisticated navigation and coordination tactics, adapting adeptly to the variable requirements of real-time warehouse scenarios. Through comprehensive experiments, we demonstrate significant enhancements in autonomous decision-making and operational efficiency, thereby validating the effectiveness of merging MARL with CL in intricate, multi-agent contexts. Our findings align with recent advancements in MARL applications for robotics 1 and underscore the potential of curriculum-based learning strategies in complex environments.