AUTOMATED GRASP PLANNING AND EXECUTION FOR REAL-WORLD OBJECTS USING COMPUTER VISION AND TACTILE PROBING

摘要

This article describes an automated grasping system suitable for complex 2.5D real-world objects (i.e., objects with height < width). The system hardware consists of a robotic manipulator, a three- fingered dexterous hand with a palm-mounted CCD camera, and a PC. The shape of the given object, along-with its position and orientation, are not known by the system a priori. A 2D model of tie object is first obtained using computer vision. This model, along with height information obtained using online tactile measurements, forms the input to the grasp planner. The planner extends our second-order limited mobility grasping theory to generate optimalimmobilizing 2D grasps. Two novel quality metrics are employed in the optimization. The grasp-planning theory is applicable to grasps with three or more fingers. The tactile information is used to extend the grasp to 2.5D objects. We assume that the fingers of the robotic hand will provide an out-of-plane constraint for the object. Experiments are performed on three complex shaped automotive parts using a three-fingered, nine-axis, dexterous hand. The total time for object modelling and grasp planning was under 0.3 s for each part. The parts were successfully grasped and immobilized in all of the tests performed.