Research On Adaptive Path Planning Method For Multi-sensor Planetary Rover In Dynamic Environment

When moving on the surface of the planet,as a key component of the autonomous decision-making system,the path planning of the planetary rover is an important guarantee for its safe and efficient scientific detection.Especially in the future,in order to detect high-value scientific targets on a large scale,there may be scenarios where astronauts and multiple mobile robots work together,which is accompanied by a more complex dynamic environment.In the traditional planning method,all the behavior of the planetary rover comes from the pre-defined rules of the ground personnel.However,the operation environment of the planetary rover is not completely known,so it is necessary for the planetary rover to have certain adaptive and self-learning capabilities for environmental changes.In order to further enhance the autonomous decision-making ability of the planetary rover,and solve the problem that traditional artificial planning framework relies too much on map information,and an end-to-end path planning method based on the deep reinforcement learning is proposed,which maps the action instructions directly from the sensor information to the planetary rover.At the same time,different neural network structures are used to process different sensor information,and finally the environmental features are merged together to form a path planning method for multisensor planetary rover based on D3 QN PER.Firstly,the basic theory of deep reinforcement learning is deeply studied.The convolutional neural network is used to process visual image information,the long-shortterm memory is used to process the point cloud information of LIDAR and self-state information,and then the fusion scheme of environmental features of the planetary rover is given.At the same time,taking the advantage of other deep reinforcement learning algorithms,the D3 QN PER algorithm is applied to the path planning system of multisensor planetary rover,and then output velocity and angular speed instructions to control its motion.Secondly,build a simulation environment and verify the effectiveness of the algorithm in three steps.The first step is to test whether the algorithm of deep reinforcement learning can guide the planetary rover to the goal point in an open environment.According to the results of path planning,D3 QN PER has more advantages than other derivative algorithms of DQN.In the second step,the gravel and rock on the surface of the planet are simplified into a static obstacle environment.The obstacle avoidance performance of the path planning method for multi-sensor planetary rover based on D3 QN PER is verified,and then compare it with the traditional RRT* algorithm.In the third step,the astronauts and multiple mobile robots on the surface of the planet are simplified into a dynamic obstacle environment,and the network model trained in the static obstacle environment is directly loaded.The experimental results show that the planetary rover has strong adaptive ability to environmental change.At the same time,the method is compared with the traditional improved artificial potential field method.Finally,by loading the network model trained in the simulation environment,the Jackal mobile robot is used to verify the practicability of the path planning method based on D3 QN PER in the real environment,thus providing reliable experimental support for the actual detection.