Decentralized Multi-robot Path Planning using Graph Neural Networks

Communication plays a key role for fruitful decentralized multi-robot path planning. However, it is quite difficult to discern which insight is necessary to perform the assignment, when and how it should be exchanged among robots. To avoid these problems and go beyond the ad hoc design of heuristics, we introduce an integrated model that generates coherent, inter-communication and decision-making for robots operating in a confined working environment. The architecture of our work includes a convolutional neural network (CNN) to achieve sufficient patterns from nearby sensing and a graph neural network (GNN) to share these characteristics within robots. This trained network mimics an expert algorithm and can be employed online in decentralized planning where we have only local interaction and observations. In the simulation-based evaluation, we steer group of robots to their goals in 2D complex work environments. We compute the success probability and total cost along each of planned strategies. The performance of our algorithm is nearly the same as our expert algorithm, which proves the potency of the advocated technique. Specifically, we demonstrate that our model allows for testing on new cases (large environments, a larger number of robots).