Information Processing And Decision Making Under Multi Robots Combat

The information processing and decision making methods of robots are the key issues in multi-robot adversarial environments.However,there is still much work to be done in the area of environmental information perception,robot state estimation and decision making at this stage.With the ICRA DJI Robo Master Artificial Intelligence Challenge as the background of this thesis,this paper mainly studies information processing and decision-making problems in a multi-robot confrontation environment,aiming to achieve robot target detection,state estimation,and autonomous decision-making.The main work of this paper are summarized as follows:First of all,the tasks and metrics of visual inspection are analyzed and the configuration scheme of the visual sensor is given;Considering the limited computing performance of the mobile device,a detection-tracking framework which is composed of two detection networks is proposed.The detection network and tracking network is designed basing on YOLOv3 and Center Net.All of these networks were trained with the dataset generated by the simulation environment.Tested with both the simulation dataset and the real-world dataset,the results show that the detection and tracking network based on YOLOv3 can achieve higher accuracy and frequency,which meet requirements of the competition.Secondly,in view of the defects of Gaussian noise assumption and instability when noise changes in the Extended Kalman filter,a Deep Adaptive Extended Kalman filter algorithm is proposed for the estimation of nonlinear,non-Gaussian systems by embedding a deep neural network into a standard Extended Kalman filter framework.With several nonlinear system simulation experiments,the results show that the proposed Deep Adaptive Extended Kalman filter has better estimation performance and stability than traditional filtering methods.Thirdly,the requirements of the strategy simulation environment on simulation frequency,shooting realism and planning realism are analyzed.the overall architecture of the multi-robot confrontation simulation environment platform is designed,including the combatzone environment,path planning,robot system and visualization.In order to accomplish high frequency path planning while making sure the gereated path is similar to the realistic path,two planning methods based on optimized cubic spline interpolation and Social GAN are proposed.The experiments show that the designed strategy simulationplatform is capable of completing the game simulations realistically and efficiently,providing strong support for the designing and training of the reinforcement learning strategy algorithms.Finally,considering the requirements of sample efficiency,scalability,and generalization capability,the decision task is decomposed into two tasks,direct mission and indirect mission.A hierarchical reinforcement learning based multi-robot decision making method is proposed.Using the MADDPG algorithm to training the attacking modules of the whole decision making algorithm,the results of the experiments show that the robot can learn the behavior of the decomposed task after a short time of training,which improve the feasibility and robustness of the whole decision algorithm.