Design and control of a nonprehensile impulse manipulator

Abstract

In modern manufacturing industries, small components such as bolts and nuts of a complex assembly are usually delivered to the plant in big loose batches. To autonomously feed those components into ongoing assembly processes with celerity, vibratory bowl feeders (VBF) were developed in 1950 to perform singulation, orientation, and manipulation tasks. In the past 20 years, however, as robot assembly systems became a prominent part of the new and more versatile manufacturing environment, VBFs appeared to be less suitable as 1) each of them is designed for one specific part only, and 2) the cost to design and tune a new variation is expensive. This thesis proposes an alternative design of a nonprehensile impulse manipulator with the corresponding control method for singulation, orientation, and manipulation by means of seven fixed-position variable-energy solenoid impulse actuators located beneath a semi-rigid part supporting surface. To supervise the manipulator, a 640p webcam with computer vision tools was included to provide part pose information. To control the device, machine learning algorithms were used to generate a part-specific control policy that bring the part to a user specified target pose. The device was tested by manipulating a six-faced craps-style die and an imprecise flat square wooden nut from a child's assembly toy. Compared with the benchmark policy, the trained optimal policy was able to flip the die to any desired face with six times higher probabilities and stand the flat nut up on its less stable pose with two times higher probabilities. The device was then put into a collaboration task with a 6-DoF robot manipulator to complete a manipulation task on the six-faced die. The resulted average execution time was faster than most state-of-the-art manipulation tactics.