Research On Navigation Of Service Robot Based On Reinforcement Learning

With the development of technology and the improvement of living standards that people pursuit, service robot as one new member of robot family has come into being, and it has played a significant role in medical assistance, family services, entertainment and other areas. And the autonomous navigation as the fundamental and important capability of service robot is the basis for other application; but the environment which service robot works in is complex, unstructured and unknown, so a higher requirement for its navigation is proposed. In recent years, with the development of artificial intelligence, how to learn by interacting with the environment independently to increase the intelligence of service robot and make it better adapt to the complex environment has become a hot research topic of robotic. And the reinforcement learning as one method of machine learning because doesn't need mathematical model and prior knowledge, just need to interact with the environment by trial and error learning can get the corresponding optimal control strategy. Therefore, it can be applied to the service robot navigation control.This paper after studies the characteristic and previous study results of reinforcement learning and combine with the character of service robot working environment, the reinforcement learning is applied to the navigation control of service robot. Firstly, for the exploring navigation control of service robot in unknown environment, propose the appropriate state space method, design reward foundation combining discrete and continuous reward, and in order to accelerate the learning speed, the Q-Learning joined eligibility trace reinforcement learning algorithm is introduced to this paper. And simulation results show that the reinforcement learning system designed is reasonable. Secondly, study the reinforcement learning applying in the path planning based on the map information of environment. We adopt discrete grid state partition method and introduce a reinforcement learning method learning on "the policy of wheel rotation" in the path planning. We prove that this method can shorten learning time and assure the convergence of learning process and the optimality of planning path by simulation experiment. Finally, make a research on the navigation of service robot in real dynamic environment. To overcome easily trapped in local minima and the poor ability to adapt to dynamic environment and other problems, a comprehensive style control strategy integrating path planning based on reinforcement learning and fuzzy obstacle avoidance is introduced In this paper. It is well absorbed the advantages of global path planning based on reinforcement learning and fast obstacle avoidance based on fuzzy. And we verify this comprehensive style navigation control can well achieve the task of navigation in dynamic environment.