Curriculum-Based Reinforcement Learning for an Effective Multi-Agent Path Planning Algorithm in Warehouse Scenarios

Multi-robot systems can provide substantial increase in efficiency and/or flexibility in different scenarios. Applications in various settings have been studied in the literature, such as disaster management, surveillance, object transportation as well as search-and-rescue. A particular case that can highly benefit from the employment of multiple agents is the logistics in a warehouse scenario. This work proposes an multi-agent Q-learning based algorithm with curriculum learning and transfer learning to perform the path planning process. With progressively more complex stages of training as well as knowledge transfer from one stage to another, the algorithm is capable of achieve high success rates. In order to validate the proposed method, simulations were done to compare it to other combinations of the used techniques, as well as using Q-learning only. Scalability tests were also performed. The proposed method achieved up to 94% success rate after training.