Intelligent robot control using ultrasonic measurements

Abstract

We present an intelligent robot control system which employs low-quality ultrasonic measurements to perform high-precision recognition and grasping tasks. The system adaptively restores the ultrasonic image using approximators based on neural networks. Neural networks are also applied to perform object classification and a grasp planning task. Since the grasp planning is not unique, we developed a novel learning scheme that uses the expectation maximization approach. The resulting system works precisely and reliably. The underlying methodology can be extended to other low-quality data problems.