Vision-based methods for state estimation and control of robotic systems with application to mobile and surgical robots

For autonomous systems that need to perceive the surrounding environment for the accomplishment of a given task, vision is a highly informative exteroceptive sensory source. When gathering information from the available sensors, in fact, the richness of visual data allows to provide a complete description of the environment, collecting geometrical and semantic information (e.g., object pose, distances, shapes, colors, lights). The huge amount of collected data allows to consider both methods exploiting the totality of the data (dense approaches), or a reduced set obtained from feature extraction procedures (sparse approaches). This manuscript presents dense and sparse vision-based methods for control and sensing of robotic systems. First, a safe navigation scheme for mobile robots, moving in unknown environments populated by obstacles, is presented. For this task, dense visual information is used to perceive the environment (i.e., detect ground plane and obstacles) and, in combination with other sensory sources, provide an estimation of the robot motion with a linear observer. On the other hand, sparse visual data are extrapolated in terms of geometric primitives, in order to implement a visual servoing control scheme satisfying proper navigation behaviours. This controller relies on visual estimated information and is designed in order to guarantee safety during navigation. In addition, redundant structures are taken into account to re-arrange the internal configuration of the robot and reduce its encumbrance when the workspace is highly cluttered. Vision-based estimation methods are relevant also in other contexts. In the field of surgical robotics, having reliable data about unmeasurable quantities is of great importance and critical at the same time. In this manuscript, we present a Kalman-based observer to estimate the 3D pose of a suturing needle held by a surgical manipulator for robot-assisted suturing. The method exploits images acquired by the endoscope of the robot platform to extrapolate relevant geometrical information and get projected measurements of the tool pose. This method has also been validated with a novel simulator designed for the da Vinci robotic platform, with the purpose to ease interfacing and employment in ideal conditions for testing and validation. The Kalman-based observers mentioned above are classical passive estimators, whose system inputs used to produce the proper estimation are theoretically arbitrary. This does not provide any possibility to actively adapt input trajectories in order to optimize specific requirements on the performance of the estimation. For this purpose, active estimation paradigm is introduced and some related strategies are presented. More specifically, a novel active sensing algorithm employing visual dense information is described for a typical Structure-from-Motion (SfM) problem. The algorithm generates an optimal estimation of a scene observed by a moving camera, while minimizing the maximum uncertainty of the estimation. This approach can be applied to any robotic platforms and has been validated with a manipulator arm equipped with a monocular camera.