Multi-Robot Obstacle-Avoidance Formation Based on Graph Neural Networks and Imitation Learning

Obstacle avoidance is the basis for multi-robot systems to perform various tasks. Considering real-world scenarios where multi-robot formation has limited perception range and communication capabilities, achieving coordinated obstacle avoidance using decentralized control method while maintaining formation stability poses a significant challenge. In this paper, we propose a decentralized, learning-based control method using graph neural networks to achieve stable control of the formation. This method improves the convolution of the graph neural networks by combining the information of obstacles, and at the same time encodes the information of obstacles as the input of the graph neural networks as well. We use a modified artificial potential field method as an expert algorithm and train graph neural networks to imitate the outputs of the expert algorithm. This ultimately achieves obstacle avoidance for multi-robot formation, which is validated on physical robots. The experimental results demonstrate that the performance of our method is close to that of the expert algorithm, with a higher success rate of obstacle avoidance and a lower number of collisions. Moreover, it is more advantageous in maintaining the formation steady state, proving the effectiveness of our method.