Dual control approach for automatic docking using monocular vision

This dissertation addresses the problem of control input design for a system whose state observability is dependent on it's state trajectory. Specifically, it describes the development of a new control algorithm which exploits the coupling between observability and trajectory to improve the overall stochastic control performance. Many control systems which suffer from a lack of observability due to a limited sensing capability can benefit from this proposed control technique.;This research was motivated by automatic docking control for an unmanned vehicle whose goal is to reach a target dock at a desired speed and direction without any human aid or intervention. Accomplishing this requires the ability to determine relative position accurately, and this in turn depends on the degree of observability of the system.;The approach used to develop this new control algorithm is to pose the problem as a nonlinear stochastic control problem incorporating a cost term which weights the system state uncertainty. The resulting form of the problem is similar to linear quadratic control and can be solved using computationally efficient methods. The overall result is a systematic controller design procedure that avoids both the computational complexity of dynamic programming approaches and the heuristics typical of approximate methods.;To demonstrate the feasibility and utility of this new approach, it is applied to the limiting-case problem of monocular vision-based docking of a kinematic vehicle. The monocular vision sensor provides bearings-only information without range, and thus the system is unobservable in terms of traditional linear observability. The newly designed docking controller generates trajectories that both satisfy the docking control goal, and improve the system's observability simultaneously. Both simulation and experimental results are presented for a wheeled rover system.