Reserach On Robot Path Planning Algorithm Based On Deep Firefighting Learning

Path planning is an important component of autonomous navigation for mobile robots and has received widespread attention and research from scholars.How to enable robots to autonomously plan and navigate to target points in complex and unknown environments is a key research topic in this field.Meanwhile,autonomous path planning and decision making without supervision and map guidance is also a current hotspot.In view of this,this paper develops the research of path planning and navigation tasks based on reinforcement learning methods.The main work is as follows.First,to address the problem that traditional algorithms are prone to fall into local optimum in complex unknown environments,we use proximal policy optimization algorithm for robot’s path planning and navigation task.An end-to-end convolutional neural network is constructed based on visual inputs to process the state inputs,output the linear and angular velocities of the robot,and a dense reward function is designed to maximize the reward by continuously interacting with the environment.The algorithm is learned and trained in a Gazebo simulation environment,and the performance of the algorithm is verified in different test environments.According to the experimental results,it is demonstrated that the robot can complete the planning and navigation tasks without collision after the algorithm converges.Secondly,an algorithm named contrastive representation with attention based on Proximal Policy Optimization(CRA-PPO)is proposed to address the problems of low efficiency of visual RGB input and low generalization of the algorithm.The algorithm trains a neural network with strong representational ability by introducing contrast learning to provide efficient state input for reinforcement learning,while embedding an attention mechanism in contrast learning to enable the network to compare features based on useful information to further strengthen the representational ability and improve the network training speed and algorithm generalization.The performance of algorithm is also verified in both simulation and real world,and the experimental results demonstrate that the convergence speed and stability of the CRA-PPO algorithm are significantly improved in both static obstacle environments and dynamic obstacle environments.Finally,for the problem that reinforcement learning algorithms fail in planning due to low exploration efficiency in complex and long sequence environments,a reinforcement learning approach based on information fusion and hybrid controller(Hybrid Controller based on Proximal Policy Optimization,HC-PPO)is proposed.The pure visual RGB image input is replaced by a state input consisting of the robot’s state,target point information,Li DAR information and visual information,while the artificial potential field method is selected as the prior controller to provide guidance for the policy controller of reinforcement learning,which plays a complementary effect.The performance of the HC-PPO algorithm is verified in simulation and real world,solving the problem of low exploration efficiency of reinforcement learning algorithms in complex,long sequence environments.