| Motion planning algorithms are the key to the autonomous operation of mobile robots such as wheeled or legged robots.Traditional algorithms mainly use optimization methods for motion planning based on information such as the robot model and environmental perception.They lack intelligence and cannot meet the requirement for intelligent robots.It is still a big challenge to build the robot’s motion planning ability from zero to a high level of intelligence.The goal of this dissertation is to adopt a reinforcement learning-based end-to-end approach to enable mobile robots with no control at all to gain intelligent control capabilities.Regarding the motion planning of wheeled mobile robots,the main research goal of this dissertation is to propose a reinforcement learning-based end-to-end intelligent controller,which uses two-dimensional laser perception of the environment and the speed,distance,and direction of pedestrians,outputs the robot’s linear and angular velocities,and considers social compliance(human walking habits),without using any map.The method in this dissertation combines imitation learning and reinforcement learning so that the training time of reinforcement learning is reduced by about 60%;implements safety constraints so that the collision rate is decreased by about 50%compared with simple reward design;adopts the strategy of cooperative obstacle avoidance so that socially compliant navigation behavior is acquired.The experiment results prove that the end-to-end controller proposed in this dissertation has high safety,feasibility,and generalization ability to new environments.Regarding the motion planning of hexapod mobile robots,the main research goal of this dissertation is to propose a reinforcement learningbased end-to-end intelligent controller,which uses internal perception such as robot height,posture,speed,joint angles and angular velocities,foot-toground contact state and outputs robot joint control without environmental perception.This dissertation proposes a data augmentation algorithm to generate training data of walking straight in different directions,expands the distribution of training samples;combines representation learning and Siamese network to learn the equivalence between different trajectories;deploys batch reinforcement learning so that the hexapod robot can learn to walk in different directions after it knows how to walk in one single direction.The experiment results prove that the end-to-end controller proposed in this dissertation has a strong target orientation ability and has a great improvement compared to other baseline controllers.The average shortest distance to the goal is improved by 6%,which verifies the feasibility of the proposed intelligent end-to-end controller. |
