Research And Implementation Of Visual Obstacle Avoidance Technology For Mobile Robot Based On Inverse Reinforcement Learning

With the continuous development and popularization of mobile robots and artificial intelligence technology,people’s demand for the intelligence of mobile robots is also increasing.As one of the basic functions of robots and vehicles,obstacle avoidance ensures safe and stable driving of robots and the like.Traditional obstacle avoidance algorithms for autonomous navigation robots often need to provide a lot of prior information,such as maps and point cloud data.However,in practical application scenarios,the above information may not necessarily be obtained due to objective conditions.For example,the scene map needs to be manually created and acquired in advance,and when the robot is in an unfamiliar environment,there is no map data available.In addition,collecting point cloud data,etc.,often requires adding expensive radar sensor equipment to the robot,which also increases the cost of obstacle avoidance.Taking common indoor scenes as an example,this thesis proposes and implements a visual obstacle avoidance technology for mobile robots based on inverse reinforcement learning.This technology can be used as a new method for robots and vehicles to avoid obstacles in unknown indoor environments.It can be applied to obstacle avoidance in many scenarios with limited costs and conditions,such as indoor environments with unknown maps,and scenes with frequently changing environments.In order to further reduce the cost of obstacle avoidance sensors,this thesis selects a low-cost visual sensor,and builds a network model according to the characteristics of visual obstacle avoidance tasks,so that the features extracted by the network take into account both the time and space dimensions.How to make the mobile robot avoid obstacles intelligently without a map is the core content of this thesis.This thesis uses the inverse reinforcement learning(Inverse Reinforcement Learning,IRL)algorithm model to learn the reward function from the expert obstacle avoidance data,and combines reinforcement learning(Reinforcement Learning)Learning,RL)algorithm model to achieve intelligent obstacle avoidance of mobile robots.In order to reduce the trial and error cost in the training process of the algorithm model,the algorithm model implemented in this thesis will be trained in the simulation environment and tested on the self-developed robot platform in the laboratory.This thesis uses ROS and Gazebo to build a simulated indoor environment for obstacle avoidance algorithm training,and builds a mobile robot kinematics and dynamics simulation model based on the robot designed by the lab.In this thesis,a framework that satisfies obstacle avoidance IRL and RL training and reasoning processes under ROS is constructed.This framework can be used in both the Gazebo simulation environment and real ROS robots,and realizes the unified management of training and reasoning processes.The test results show that the method implemented in this thesis has a good obstacle avoidance success rate in both the simulated environment and the real environment(71%in the simulated environment and 65%in the real environment).Better obstacle avoidance effect,which can realize indoor obstacle avoidance of mobile robots without a map and only use visual sensors,which has certain practicability.