Learning from Virtual Experience: Mapless Navigation with Neuro-Fuzzy Intelligence

摘要

Traditional robot navigation approaches normally rely on creating a precise map of the environment which is a computationally expensive procedure and highly depends on an accurate sensory system. Even for motion planning in similar terrains, the planner needs to prepare or obtain a map beforehand. In this paper, this issue is addressed, and a neurofuzzy motion planner is presented for mobile robot navigation without a map. We show that, by means of a virtual experience model and a neuro-fuzzy system, a mapless motion planning approach can learn basic navigation primitives in simple obstacle arrangements without any prior demonstration. The virtual experience model creates a large number of test environments with a random set of arbitrarily shaped obstacles and places the robot in a random pose with different start and goal positions in different instances. Then, based on the readings of the robot's sensors and a collection of predefined general linguistic rules, a set of control commands including the robot's linear and angular velocity is calculated as the outputs of the virtual experience. The resulting dataset is then loaded into an adaptive neuro-fuzzy inference system to create and optimize a fuzzy motion planner using the subtractive clustering method and a hybrid technique combining the back-propagation algorithm and the least square adaptation method respectively, which guides the robot in simple unknown environments without requiring a global obstacle map. To validate the effectiveness of the proposed model, the motion planner was implemented on a nonholonomic differential drive robot to test its performance in two real navigation tasks. Experimental studies show that the proposed mapless motion planner can efficiently guide the robot in similar arrangements of convex obstacles.