| Deep reinforcement learning,with its dual advantages of deep learning and reinforcement learning,has not only made significant theoretical breakthroughs,but has also gained practical applications.With the widespread application of robots,the working environment is more complex and changeable,the requirements for robot navigation and intelligence become more demanding,and traditional path planning methods no longer meet the needs of modern robots.At this time,it is necessary to further improve the robot’s adaptive ability and autonomous decision-making ability.Deep reinforcement learning methods in artificial intelligence have shown excellent performance for robot path planning,which is achieved by interacting with the environment through trial-and-error.In this paper,the deep reinforcement learning method is used as an algorithm for robot path navigation and is studied as follows.First,the underlying theory of deep reinforcement learning is studied in depth.Aiming at the random sampling mechanism in the deep reinforcement learning method to sample samples with equal probability,leading to the possibility of over-using samples with low information in the training process of the network,a subsampling sampling mechanism is proposed to increase the frequency of playback of more informative samples according to the TD_error distribution of the samples.The subsampling sampling mechanism is applied to deep Q network to constitute a deep Q network method based on subsampling sampling,and the method is validated in a Mountain Car scene.Experiments show that the method can increase the convergence rate of the algorithm and improve the learning effect.Secondly,in order to make the deep reinforcement learning method adopted in this paper have better convergence and planning capability in the robot path planning task.The information of obstacles near the robot and the orientation information of the target location obtained through sensors during the robot’s movement are discretized,and then the applicable robot environment model and action space are designed.A nonlinear segmented reward function is designed to provide timely feedback for each action of the robot’s taken.The exploration and development strategy in deep reinforcement learning is improved by using a combination of greedy strategy and Boltzmann distribution function to reduce invalid exploration and improve the stability of the algorithm training process.Finally,a 3D robot simulation environment based on ROS and Gazebo is built to test the path planning capability and generalization of the improved method in an unknown environment,as well as the effectiveness of the path planning framework design.Firstly,the robot model and simulation environment are built based on the designed robot path planning method.Secondly,the improved method is tested in the simulation map to see if it can effectively avoid obstacles and successfully guide the robot to the target location to obtain the best or second best path.Finally,load the network model trained by the improved method SS-DQN-GB,and test the generalization and adaptability of the method on 16 maps with different environmental information. |
