Online Path Planning For Mobile Robots In Unknown Indoor Scences

The mobile robot is a technology that includes the latest research results of the automatic control,artificial intelligence,mechanical design,computer science,etc.It has rapid development in these years.The mobile robot is an intelligent mechanism that combines the ability of perception,decision-making and movement.It needs the sensor to receive the sensory information in real time,and then through the signal processing,finally gives the decision-making action.Among them,the path planning navigation problem is an important part in the research of mobile robot.The traditional path planning algorithm has a history of more than half a century.The basic planning process is that the environment and the robot mode has been given,then formulate the target points to be planned,finally automatically calculate the motion path of the robot.However,these planning algorithms have two prerequisites: global maps are known and target points are known,but our ideal mobile robots need to have generalization capabilities for autonomous navigation in unknown new scenes,as well as collision-free autonomous actions that can have no single target point in complex scenes.Therefore,how to realize the online path planning of mobile robots in unknown indoor scenes has become our main research direction.As a branch of machine learning,reinforcement learning has been around for decades,but it has not been widely used due to its inherent limitations in computational efficiency and input feature representation.Along with the rapid development of deep learning technology in recent years,deep intensive learning technology combined with deep learning began to enter people's field of vision after AlphaGo defeated the world championship in 2016,and has al ready tried and applied in many scientific research fields.Based on mobile robot path planning technology and deep reinforcement learning technology,we propose a new algorithm named adaptive-attention DQN with,which enables mobile robots to explore maps independently in an unknown indoor dynamic environment and real-time obstacle avoidance,And solve the problem that the robot enters the dead end multiple times due to the lack of memory when the scene encounters a dead end road.At the same time,after training multiple rounds to achieve convergence,the robot is placed in a new unknown scene,it also has better path planning and obstacle avoidance capabilities.The simulation results show that the proposed method of adding the adaptive attention mechanism to the deep reinforcement learning algorithm has a better performance than the latest algorithm for deep reinforcement learning based on value iteration(Rainbow DQN).