Research On Deep Learning Technology Of Unmanned System Supporting Group Motion Planning

The application mode of intelligent unmanned system is changing from scattered deployment,single task to large-scale applications and intelligent cooperation.Group motion planning is one of the core issues to realize the transformation of this application mode.The traditional "human-designed" cooperative method is based on the predictability and unchanging assumption of task objectives and the environment.It needs to obtain comprehensive and specific information of tasks and environment in advance.The corresponding system is only limited to the tasks specially designed for it,and can not adapt to the situation beyond program design or(machine learning)training.On the contrary,the situation in the real environment is changing rapidly,all kinds of predictable and unexpected factors may appear,and the mission objectives at some levels may also change,so it is difficult for the pre-designed group intelligence of the unmanned system to deal with.Deep learning technology combines low-level features to form more abstract and high-level representations attributing to the internal characteristics of data based on its multi-layer neural network architecture,massive data input and more efficient hardware support.In this paper,the research object is multi-agent motion planning;the basic method is deep learning technology;the aim is obstacle avoidance planning of intelligent unmanned system,the learning of cooperative ability to deal with dynamic and noisy environments.In this paper,firstly,a group path planning method based on the complementary advantages of heterogeneous robots is proposed,which uses feature fusion technology to fuse the views of multiple robots to plan the paths;then,a multiagent cooperative dynamic environment exploration method based on deep reinforcement learning technology is proposed to cope with the changes of dynamic environment;finally,a multi-agent fault tolerance method based on multi-head attention mechanism is proposed.The fault-tolerant method enables multiple agents to distinguish useful information in extremely noisy environments and accomplish tasks cooperatively.The main work of this paper is shown as follows:(1)A path planning method based on multi-robot visual feature fusion technology is designed.In view of different ability and different angle of view,this paper proposes a path planning method based on the complementary advantages of heterogeneous robots.On this basis,the specific task of multi-robot trail following is selected,and the path planning method based on multi-robot visual feature fusion is realized,which improves the robustness of the system.(2)A cooperative dynamic environment exploration method based on deep reinforcement learning is designed.Aiming at the problem that traditional "humandesigned" cooperative method is difficult to deal with various unexpected factors effectively,this paper proposes to use "learning-based" methods to train multiple robots to master cooperative strategies,without obtaining comprehensive and specific information of the task and environment in advance.At the same time,attention mechanism is introduced to help the robots choose valuable message while communicating,so as to effectively deal with the interference of dynamic obstacles.(3)The fault tolerance mechanism to deal with the noises in the process of multi robot cooperative exploration is designed.In order to solve the problem that it is difficult to learn effective cooperative strategies in the process of multi robots' exploration tasks in extremely noisy environments,this paper proposes a multi-head attention mechanism to help robots select correct and valuable information in the process of communication,and then correctly complete the process of state value estimation to deal with the noisy environments.Based on the above implementation scheme and mechanism,experiments are carried out in open datasets and real scenes to verify the proposed methods comprehensively.Based on the above implementation scheme and mechanism,experiments are carried out in multi-robot cooperative exploration simulation environments and real environments,which proves the effectiveness and advantages of the proposed methods.