Evolutionary Neural Network-Based Obstacle Avoidance for a Planetary Exploration Rover

During space missions, a planetary exploration rover is subject to two main communication issues that are inherent to the distance and orbital difference between Earth and the target planet, the communication transmission delay and the limited transmission window.Besides, the computational resources are restricted onboard the vehicles for space missions.Based on these facts, the rover is desired to traverse on the target planet efficiently with fewer human commands and less computational needs, as well as with an autonomous decision-making scheme -allowing the rover to perform autonomous science.The proposed autonomous path planning system presents a set of genetically evolved neural network controllers for local path planning of a mobile robot.Travelling environments are partitioned into three categories according to distribution patterns of targets and obstacles.Each evolved network is adopted to direct the rover travelling in one category of partitioned environments achieving a sequence of targets with obstacle avoidance.With a set of pre-learned networks, the rover would be adaptable to traverse in new environments of specific category.Input to the network is the range and bearing data measured from current position of the rover to surrounding obstacles and the approaching target.It outputs a turning angle ratio processed to be direction of the rover for next move.Genetic algorithm is used to obtain the evolved network by developing behavior strategies through evolutionary iterations.Simulation results indicate that evolved neural controller can adapt to novel environments and generate satisfying path for the rover in a computationally economic manner.