Vision-Based Predictive Robotic Tracking of a Moving Target

Abstract

This work represents a more general approach to robotic system design than one based on predefined responses in a controlled environment. An implementation of a vision-based robotic tracking system is presented in which target trajectory predictions enable the robot to track and intercept a moving target. A host microcomputer receives target position information from a vision module, predicts the target's trajectory, and issues tracking commands to the robot controller. Five predictive algorithms are derived for implementation in the system, including a Kalman and an augmented Kalman filter. The use of one-step as well as absolute and relative n-step predictions is investigated. The "best predictor" algorithm is presented, by which one of the five predictions is selected to be used as the robotic tracking command. Using data from experimental trials, predictor results are compared and rcjbotic tracking performance and interception success are evaluated for the target both moving and after it comes to rest. Constraints limiting the applicability of this implementation are discussed and possible improvements and extensions suggested.