Research Of Multi-Robot Collision Avoidance Based On Machine Learning

Problem of multi-robot collision avoidance is an important research direction in multi-robot systems research. It plays a positive improving role in the development of multi-robot systems and robot technology. Because of the inadequacy of traditional methods (such as rule-based, behavioral approach, etc) to solve multi-robot collision avoidance problem, the research of multi-robot collision avoidance through learning becomes focus of attention. This paper focuses on multi-robot collision avoidance planning to analyze and study. Reinforcement learning algorithm is used to solve the problem, the aim is to make the robot quickly and effectively avoid all kinds of obstacles and another robots in the process of movement towards the target.Firstly, several basic theories of machine learning have been introduced, mainly including reinforcement learning algorithms and artificial neural network. Then the multi-robot system and some of the relevant knowledge has also been introduced.Secondly, with the lack of the traditional reinforcement learning algorithm in practical application, we propose a method of combining the BP neural network and Q-learning, using BP neural network to solve the mapping problem between the state space and action space in reinforcement learning. The state space reinforcement learning is used as the input of BP neural network. The output of BP neural network is corresponding to the Q value for the action. This approach not only solves the reinforcement learning in continuous state space in the application, but also improves the learning convergence speed.Then, the combined algorithm is applied to multi-robot collision avoidance problem. To simplify the input state space, we divide the multi-robot collision avoidance behavior learning into two layers based on hierarchical thinking: static obstacle avoidance behavior and dynamic obstacle avoidance learning layer. Each layer has separately been applied research in reinforcement learning methods. Meanwhile this paper designs strategies to achieve collision avoidance between robots.Finally, a simulation platform for multi-robot system has been built to verify the validity of the algorithm. Simulation results show that the research of method is better to solve the multi-robot collision avoidance problem.