Real-time adaptive control of a robotic manipulator using three-dimensional stereoscopic vision feedback
John Rasure · 1985
This research proves that three-dimensional stereoscopic vision can be used as a feedback sensor to an adaptive control system to allow a robotic arm to track a simple object. This work was completed using off-the-shelf equipment and has immediate applications in industry. The tools necessary to analyze adaptive control with vision feedback were built in the simulation phase of the research. The implementation phase of the research validated simulation techniques and also made evident some of the limitations of robot manipulator visual navigation. The vision system and control system were mathematically modeled and then simulated with C language computer programs. Modeling the geometry of 3-D stereoscopic vision was a three step process: first, a camera model was developed in the form of a homogeneous transform matrix to map object coordinates to image coordinates, then this model was used to design optimal stereo camera configurations, and finally a three-dimensional vision model was developed to efficiently map the coordinates of two images to a 3-D object location. The control system model consisted of an adaptive filter in a lattice configuration coupled with a two point extrapolater. A three-level hierarchical control system is used where the adaptive predictor resides on top and an existing planner and joint servos provide the lower two levels of control. The input to the adaptive filter and extrapolater is a sequence of object locations provided by the vision system and the output is a predicted position for the robot. The implementation phase of the research interfaced the vision/control system to an existing robot arm. A stereoscopic vision system with better than 1.0 centimeter spatial and depth resolution was built and mounted on a robot arm. Because of path planning limitations inherent in the robot used the control algorithm was simplified to a two point extrapolater combined with a bandpass filter. This tracking system coupled with vision feedback is capable of tracking a simple object at 26.0 millimeters per second and will converge to within 1.0 centimeter.