Navigational vision

Traditionally, a robot's visual system is assigned the task of reconstructing the shape of the surrounding scene in the form of a depth map, which can be exploited to solve navigation problems by means of trajectory planning, control of mechanisms, etc. Unfortunately, as an overview of the state of the art in visual reconstruction reveals, it is still impossible to reliably compute depth maps, due to the fact that all shape-from-x problems are mathematically ill-posed (they have no unique solution and/or the solution is unstable). Furthermore, judging by progress made in recent years in problems such as road following and visual servoing, it appears that the depth map may after all not be the most suitable data representation for such tasks. We propose an image-based approach to the navigation problem, in which visual processes are closely and actively integrated with robot control, and show that task-specific visual information with a minimum of structure is sufficient to accomplish classical navigational tasks such as obstacle avoidance and hand/eye coordination. Through these examples we demonstrate that the use of vision actually simplifies the solution of robotics problems, allowing real-time control without the complex calibration procedures traditionally required. It has been assumed in the past that the robot's trajectory is determined exclusively by the location of the goal and that the appropriate visual data can be acquired as the robot progresses toward the goal. In particular, the trajectory planning module never takes into account the needs of the visual module. It has been recognized that the stability and accuracy of visual algorithms are affected by the motion of the observer, but not much work has been done on deciding what constitutes a "good" motion. In the final part of this thesis, we propose a quantitative criterion for the evaluation of the goodness of a particular action and show how this additional layer of planning can be incorporated into the low-level control strategy, together with higher-level symbolic reasoning.