Learning feed-forward control for vision-guided robotics

摘要

Robots are expected to do increasingly challenging tasks in complex environments. Because of the complexity, robots must use their sensors to understand the nature of a task and to monitor its execution. Visual servoing proposes several methods for using visual feedback to control the execution of tasks. A drawback of visual servoing is that it controls the motion of a robot at the velocity level. Thus, visual servoing is neither compatible with robots that require position commands nor do they provide a natural interface for path planning in special Euclidean space. This thesis proposes an image-based feed-forward control system that is based on a formulation that is similar to its feedback counterparts. Specifically, the system requires image-based errors that are derived from the image space. Instead of mapping the errors to velocities, however, the errors are mapped to variations of the robot's position through an interaction matrix called the visuomotor function. After deriving the functions of various tasks, this thesis proposes a method to approximate the parameters of the visuomotor systems. The approximation is composed of two stages that are computed on-line. In the first stage, the parameters specific to a 3D point are estimated with an incremental least squares algorithm based on QR factorization. Then, these parameters are made available to neighboring 3D points by generalizing them in the stereo space. Experiments were done in simulations and on real robots for the following tasks: gaze control of a robotic head, 3D translations, mobile robots and 6 degrees of freedom motion. Results were also compared with a feed-forward system based on 3D reconstructions from stereo images.