Visual control of robotic tasks
Billibon H. Yoshimi · 1996
Humans are adept at eye-hand control. Robots are not. Even after many hours of laborious programming, a robot's skill is only passable. The goal of this thesis is to explore how vision can be used to give robot systems the compliance and ability necessary to perform many eye-hand coordination tasks. The thesis will describe the development of three robot systems which perform a variety of tasks requiring visual control. The first system addresses the role of vision in tracking and grasping tasks. Since humans find little difficulty in visually tracking a target and then effecting a grasp, we determined that the tracking and grasping task was an appropriate vehicle to investigate the application of vision to robot control. The experimental system is composed of a real-time 3D motion stereo tracker, a robot-gripper system capable of real-time tracking, and a set of grasping strategies. The second system investigates how visual input can be used to control an uncalibrated eye-hand system. This system performs an alignment task by exploiting a simple observable geometric effect, rotational invariance. Combined with image Jacobian-based control, this system demonstrates that uncalibrated eye-hand control can be used in situations where calibration is not available. The final system re-examines the grasping and manipulation task. This system exploits closed-loop visual feedback to control a robot and gripper as they perform a complex task requiring dextrous manipulation. The goal of this research program is to examine several fundamental questions surrounding visual control. We will explore several different ways to perform visual control and find out where and when they can be used. We will examine the suitability of vision for real-time control. We hope to show visual control is a viable option for performing real world tasks which require sensing in structured and unstructured environments.