Optimal Multi-agent Reinforcement Learning for Efficient Partially Observable Multi-robot Collaboration in Warehousing

Efficient collaboration of autonomous mobile robots or multi-robots in order fulfillment tasks in warehouses enhances the delivery rate. However, the limited visibility of the working environment and varied task configurations can affect their throughput, i.e., delivery rate. Multi-agent reinforcement learning (MARL) and a MARL environment with partial observability replicating real-world warehouse order fulfillment tasks, in this backdrop, become a natural choice for the optimization of autonomous multi-robot collaboration under partial observability. To maximize the delivery rate through efficient multi-robot collaboration, this research proposes Multi-Agent Self-attention Recurrent Q-Learning (MASRQL), which is enabled by a general-purpose Q-function approximator, Deep Self-attention Recurrent Q-Network (DSRQN) and harnesses self-attention and recurrence mechanisms. The findings of the empirical analysis corroborate that the approach enhances efficiency by achieving 15.4%, 29%, and 26.2% more average delivery rate than the closest performing baselines in the respective task configurations, showcasing adaptability to the varying task configurations.