| As an important branch of intelligent robots,mobile robots have always been the research hotspots of many scholars,and their path planning and obstacle avoidance are also the core contents of mobile robot technology research.In recent years,with the extension of mobile robot application scenarios,the working environment of mobile robots has become more and more complex,which puts forward higher requirements for mobile robots.At present,most of the path planning algorithms are obtained on the basis of the environment model.It is difficult for mobile robots to plan paths autonomously in unknown environments,and they do not have adaptability to the environment.Therefore,combined with artificial intelligence theory,this paper proposes a path planning model based on deep reinforcement learning,which overcomes the shortcomings of high model dependence and low autonomy in traditional algorithms.In view of the low utilization rate of experience samples in the training process of deep reinforcement learning,a comprehensive experience playback mechanism is designed.At the same time,the Probabilistic Roadmaps(PRM)algorithm is integrated to solve the problem that the deep reinforcement learning path planning algorithm has poor path planning ability in large scenarios.The main research of this paper is as follows:Firstly,the research status of path planning and reinforcement learning is analyzed,and the background knowledge of reinforcement learning and Deep Deterministic Policy Gradient(DDPG)is described in detail.In view of the fact that the DDPG algorithm does not consider the priority of the samples,and some important information with high priority cannot be sampled in time,resulting in low sampling efficiency and slow training speed,comprehensive experience playback is used for improvement.Validated on the Gym platform.This method reduces the training time and improves the learning ability of the system.Secondly,in order to introduce the reinforcement learning method into the path planning of mobile robots,an autonomous path planning algorithm framework for mobile robots based on DDPG-APER is built.Turtlebot3 is selected as the experimental robot,and the kinematics model is designed.Combined with the task essence of path planning,the state space,action space,reward function and overall process of the robot are determined.At the same time,a certain fully connected layer and a suitable activation function are used to build two deep network models,Actor and Critic,so that the network model can better deal with highdimensional data features.Finally,the effectiveness of the path planning algorithm proposed in this paper is verified by simulation experiments.The 3D simulation was carried out using the ROS platform and Gazebo.A mobile robot simulation model and an indoor dynamic experimental scene were built,and a comparative experiment was set up.In the short-range path planning scenario,the path planning algorithm based on DDPG-APER can better achieve local planning in unknown environments,and the path-finding efficiency and path-finding length are greatly improved.Aiming at the shortcomings of DDPG-APER in the remote path planning scenario,a combined planning algorithm of PRM+DDPG-APER is designed,which discretizes the space and realizes the remote path planning task in multiple segments,which effectively improves the generalization ability of the reinforcement learning algorithm.,which improves the effect of path planning. |
