Feature localization using neural networks for cleaning robots with ultrasonic sensors

A feature localization method for simple mobile robots like domestic vacuum cleaning robots is proposed. Here, the proposed robot system has two wheels and four ultrasonic transmitter and receiver pairs. The purpose of the sensor system is using as small number of sensors as possible whereas the robot avoids collision and gathers information of the environment. To avoid collision with a small number of sensors, each sensor must have wide-angle sensibility; however, such a wide-angle characteristic leads to ambiguity of bearing angle information and spoils high-level applications like localization. To cope with this limitation, a neural network system is proposed which extracts the bearing angle of the concave corner feature utilizing distance and amplitude information of the sensors. Under the constraint of the bearing angle, relative location of the feature against the robot is found in some error bound. By comparing the feature location with a priori map, the robot would be able to correct its pose. Experimental results with real robot and ultrasonic sensors are provided.