Active Perception for Autonomous Systems : In a Deep Space Navigation Scenario (Aktive Wahrnehmung für autonome Systeme : In einem Weltraumsnavigationsszenario)

Autonomous systems typically pursue certain goals for an extended amount of time in a self-sustainable fashion. To this end, they are equipped with a set of sensors and actuators to perceive certain aspects of the world and thereupon manipulate it in accordance with some given goals. This kind of interaction can be thought of as a closed loop in which a perceive-reason-act process takes place. The bi-directional interface between an autonomous system and the outer world is then given by a sequence of imperfect observations of the world and corresponding controls which are as well imperfectly actuated. To be able to reason in such a setting, it is customary for an autonomous system to maintain a probabilistic state estimate. The quality of the estimate – or its uncertainty – is, in turn, dependent on the information acquired within the perceive-reason-act loop described above. Hence, this thesis strives to investigate the question of how to actively steer such a process in order to maximize the quality of the state estimate. The question will be approached by introducing different probabilistic state estimation schemes jointly working on a manifold-based encapsuled state representation. On top of the resultant state estimate different active perception approaches are introduced, which determine optimal actions with respect to uncertainty minimization. The informational value of the particular actions is given by the expected impact of measurements on the uncertainty. The latter can be obtained by different direct and indirect measures, which will be introduced and discussed. The active perception schemes for autonomous systems will be investigated with a focus on two specific deep space navigation scenarios deduced from a potential mining mission to the main asteroid belt. In the first scenario, active perception strategies are proposed, which foster the correctional value of the sensor information acquired within a heliocentric navigation approach. Here, the expected impact of measurements is directly estimated, thus omitting counterfactual updates of the state based on hypothetical actions. Numerical evaluations of this scenario show that active perception is beneficial, i.e., the quality of the state estimate is increased. In addition, it is shown that the more uncertain a state estimate is, the more the value of active perception increases. In the second scenario, active autonomous deep space navigation in the vicinity of asteroids is investigated. A trajectory and a map are jointly estimated by a Graph SLAM algorithm based on measurements of a 3D Flash-LiDAR. The active perception strategy seeks to trade-off the exploration of the asteroid against the localization performance. To this end, trajectories are generated as well as evaluated in a novel twofold approach specifically tailored to the scenario. Finally, the position uncertainty can be extracted from the graph structure and subsequently be used to dynamically control the trade-off between localization and exploration. In a numerical evaluation, it is shown that the localization performance of the Graph SLAM approach to navigation in the vicinity of asteroids is generally high. Furthermore, the active perception strategy is able to trade-off between localization performance and the degree of exploration of the asteroid. Finally, when the latter process is dynamically controlled, based on the current localization uncertainty, a