| The replacement of manual work by interior decoration robots is one of the main trends in the future development of the decoration service industry.The advantage of robot work is to reduce the workload of labor and ensure the quality of decoration.However,at this stage,decoration robots still need human traction to move to the work position.In order to truly realize the systematization and automation of the decoration process,it is particularly important for the robot to have the ability to navigate autonomously.As one of the main methods to achieve autonomous robot navigation,deep reinforcement learning has the advantage that the training process does not require human intervention,and it relies on the sensors carried by the robot itself to master the navigation strategy through continuous interaction with the environment,which is an end-to-end learning method.However,there are still problems that need to be solved urgently in complex and changeable decoration scenarios,mainly including: low data utilization and low algorithm performance;algorithms are difficult to migrate to other environments,and algorithms have poor generalization.This paper conducts research on the above two types of problems,and the main work contents are as follows:(1)First of all,due to the lack of open-source interior decoration simulation scenes,this paper builds an interior decoration simulation environment based on ROS and Gazebo framework;builds a turtlebot3 robot simulation model with 2D lidar in Gazebo,and designs the robot state space and action space for the follow-up algorithm Authentication provides the underlying platform.In this paper,a fully connected neural network is used to process the state information of the robot,and output the action information and state value function of reinforcement learning,and solve the above two types of problems on the basis of the proximal policy optimization algorithm(PPO).(2)Secondly,aiming at the problems of low data utilization and low algorithm performance,a navigation algorithm based on potential energy and curiosity mechanism(PPOEI)is proposed,starting from the interaction process between the robot and the environment,focusing on the study of reward signals in deep reinforcement learning navigation The importance of the task,through the idea of reward division,divides the reward information into external rewards and internal rewards.External rewards enable the robot to quickly grasp the mapping relationship between state-action-reward through artificially imposed constraints;internal rewards keep the robot interested in exploring through the difference between the predicted state and the real state,and optimize the algorithm process under the suboptimal strategy.To solve the problem of early convergence,improve the sample utilization rate of the algorithm and improve the performance of the algorithm through the idea of reward densification.(3)Finally,aiming at the problem of poor generalization of the algorithm,the idea of algorithm modularization is proposed in the way of robot self-guidance,that is,the transferable navigation algorithm of tutor-student mechanism(TSPPO),which divides the internal environment of the robot into tutor module and student module,through pretraining,the tutor module can master prior navigation knowledge,and optimize the general fixed response strategy by means of action information guidance,so that students can follow the tutor to quickly master the navigation ability;The loss function of the algorithm ensures the transferability of the algorithm,and adds dynamic decay weights to the suggested actions output by the tutor module.Experiments show that the transferable navigation algorithm based on the tutor-student mechanism exhibits good generalization in environments with associations. |
