Attention for the Allocation of Tasks in Multi-Agent Pickup and Delivery

We consider the problem of Multi-Agent Pickup and Delivery, where agents need to execute pickup and delivery tasks while preventing collisions with other agents. This is an important problem that has a significant impact on several application scenarios, particularly for logistic operations in warehouses. In contrast with previous approaches, we propose a novel attention-based deep-learning model trained to assign tasks to agents such that the resulting assignment minimises the total travel distance of agents while avoiding conflicts. Crucially, by training our model with reinforcement learning, we aim to overcome the greedy decisions that characterised previous approaches. Through an extensive empirical analysis, we compare our approach against various baselines, effectively showcasing the significant improvements achieved by our method.