Learning of Fuzzy Rules for Avoidance of a Moving Obstacle in a Mobile Robot

To coexist with human, a robot has to avoid obstacles based on human-like flexible decisionmaking. In this article, we recorded the angle and speed when a human operates a robot to avoid a moving obstacle on a developed computer simulator. Using obtained data, fuzzy rules to decide the moving direction and speed at every moment were derived as follows: as input variables, distance to obstacle, angle to obstacle, speed of obstacles, and moving direction of obstacle, were adopted. As output variables, steering angle and moving speed of robot were adopted, where it is noted not only angle but also speed is considered compared to other prior researches. Based on fuzzy-neural networks method, two networks having 4 inputs and 1 output were prepared. A membership function of input variable has 5 isosceles triangles. Fuzzy rules, number of which is 625 (=5), were assumed. Optimal center and width of each triangle were obtained so as that the network reproduces the trajectories of simulation experiment with minimum errors. The proposed method based on obtained fuzzy rules was compared with the conventional potential method and reinforcement trajectory learning method. The robot avoided flexibly and smoothly a moving obstacle like human with both short mileage and small crash rate by using proposed method on the simulator.