Path Planning For Mobile Robot Based On Deep Reinforcement Learning

With the continuous improvement of technology and the growing number of robot tasks in recent years,The environment where robots locate has become more and more complex.It's more necessary to improve the precision of robot navigation and location.Thus,it needs robots to simulate human autonomous perception of unknown environment and to make action decisions automatically,which enables them to reach their destinations autonomously and safely.An improved deep double Q network algorithm(Improved Dueling Double Deep Q-Network,IDDDQN)based on dueling network structure is proposed to solve the problem of slow convergence under the basic deep Q-Network while the robot explores the complex and unknown environment.Robots can successfully avoid obstacles and ensure safety,robustness and adaptability in complex environment when they do path planning.The main contents of this paper are as follows:Firstly,in order to make the simulation environment of the mobile robot closer to the real environment,according to the motion model,the Turtlebot of ROS system is used to simulate the mobile robot,and the laser range finder sensor is added to the model to obtain the position and environment information of the mobile robot.Gazebo is used to build the simulation environment of various types of obstacles,to achieve good path planning.Secondly,in order to solve the problem of slow convergence of robot in the process of network learning,a deep reinforcement learning network model is proposed,which combines the competitive network structure with the basic network of DDQN algorithm.The method establishes an end-to-end learning mode,using the data obtained by laser sensor as input and the Q value of robot steering action as output.In the process of network training and learning,the improved strategy of combining Boltzmann distribution with?-greedy is proposed,which can balance the exploration and use of mobile robot in action selection,reasonably select actions,and solve the problem that the robot is easily trapped in the local optimum.These conclusions are verified in the simulation experiment.Thirdly,in order to solve the problem that the mobile robot has no prior knowledge and the sample is lack of diversity,an improved resample experience replay memory is proposed,which makes full use of the existing experience to accelerate the learning speed of the mobile robot network,avoid the problem of over-estimation in training,and improve the learning efficiency.These conclusions are verified in the simulation experiment.Finally,in order to avoid the possibility of over-fitting in a training environment,IDDDQN method is verified in a simple environment,and then the robot is put into several different and complex environments to do path planning again.The simulation results show that IDDDQN algorithm can alleviate the problem of over-fitting remarkably.Simulation results show that IDDDQN method is easier to obtain the optimal path than other improved algorithms,and can explore the unknown environment better.