Multi-Robot Motion Planning for Dynamic Environment based on Deep Reinforcement Learning

In this paper, multi-robot navigation based on the deep reinforcement learning (DRL) was proposed, which could be used to solve collision avoidance between robots. Firstly, a novel neural network architecture is proposed based on a bidirectional loop module, which could map the status of obstacles to corresponding actions directly. By this way, the complexity of the calculation and decision-making process could be avoided. Then, a new reward function is proposed to improve the performance of reinforcement learning and reduce the probability of collisions in crowded environments. Subsequently, one simulation environment was built to train the proposed strategy. Also, the proposed strategy was updated and optimized by the DRL algorithm based on the actor-critic model. Experimental results show optimization results in the fields of success rate, total travel time, and average speed compared with existing methods with learning-based and optimal collision avoidance strategies.