Reinforcement Learning-based Dynamic Path Planning For Mobile Robots

In recent years,with the development of robotics,path planning and multi-robot systems have gradually become the key research in robotics.Path planning of mobile robots is usually divided into two categories: point to point path planning and complete coverage path planning.After years of development,many algorithms for point to point path planning in static environment have been proposed,while dynamic path planning is immature due to moving obstacles and some uncertain factors of the environmental information,and there are many problems need to be solved.On the other hand,the complete coverage path planning of single robot has been developed for many years and some algorithms have been presented.Because task allocation and path planning of multi-robot systems have not be solved well,the research on complete coverage path planning of multi-robot systems has many problems need to be solved.Reinforcement learning is efficient in solving uncertain sequential decision problems and has become one of the key research issues in machine learning and artificial intelligence.Meanwhile,due to the rare dependency of accurate environment model,reinforcement learning has been widely used in mobile robots' path planning.This paper focuses the research on dynamic path planning of mobile robots based on reinforcement learning.Moreover,the dynamic path planning method is combined with other methods to solve the complete converge path planning of multi-robot systems.The contributions are listed below:1.A neural network based improved Q-learning algorithm(NIQL)is proposed.The effectiveness of the NIQL algorithm is verified by comparing it with other different improved Q-learning algorithms in converage speed.In addition,by combing the NIQL algorithm and the three cubic B spline curve,a new path planning method for dynamic environment is proposed.The simulations are carried out in different dynamic environments and the results show the effectiveness of the proposed method.2.By introducing the Q-learning algorithm into the traditional RRT algorithm,proposing a new node extending function and a random node generating function,an improved RRT method(QRRT)is proposed.From simulations,the better performance of the QRRT method was certified in terms of time taken,number of expanding node and path length comparison with other improved RRT methods.Moreover,a dynamic path planning method is proposed based on QRRT,sliding window and three cubic B spline curve.The efficiency of the proposed method has been verified by simulating the process of overtaking.3.A new method based on IGA,LSPI and dynamic path planning is proposed to solve complete coverage path planning problem of multi-robot systems.In the planning method,IGA is used for task assignment of multi-robot systems and LSPI is applied for local obstacle avoidance planning.The collision between robots solved by setting the priority for each robot and the dynamic path planning based on QRRT algorithm.The effectiveness of the planning method has been verified in simple and complex environments of different conditions,respectively.