Uncertainty aware localization for autonomous robots

Autonomous mobile robots are undergoing an impressive growth. They are successfully used in many different contexts ranging from service robots to autonomous vehicles. These robots are expected to move inside the environment and, in general, to perform some operation autonomously. Their reliability strongly depends on their capability to accommodate the uncertainty generated by their interaction with the physical world. The core functionality for every autonomous mobile robots is the ability to navigate autonomously inside a known environment. The navigation task can be decomposed in identify where to go, plan and follow the route to reach the goal. In order to follow the planned path the robot needs to accommodate the actuation noise. To accommodate these noise the knowledge of the pose and speed of the robot inside the environment is needed. The more accurate the localization of the robot, the better the actuation error can be compensated for. Localisation is the process of establishing the correspondence between a given map coordinate system and the robot local coordinate system relying on its perceptions of the environment and its motion. Sensors are affected by noise, and in time, ego-motion estimation alone diverges from the robot's true pose. Robot exteroceptive sensors can give fundamental information to reset the pose uncertainty and relocalise the robot inside the environment, hence mitigating the dead-reckoning process. Most of the localization systems presented in the state-of-the-art focus on the maximization of the localization accuracy by leveraging the natural features of the environment. In these systems, the maximum achievable accuracy is tightly coupled with the perceivable information embedded in the different regions of the environment. Therefore, the localization uncertainty cannot be adapted to the level of accuracy desired by the users and only few approaches can provide guarantees on the localization performance. In contrast, by infrastructuring the environment, it is possible to obtain a desired level of uncertainty. Current approaches tend to over-design the infrastructure in dimension and supported measurement frequency. They provide far more accuracy than required in most areas of the environment in order to guarantee the tightest constraints that often are required only in limited regions. The ability to adapt to the location-dependent uncertainty is more than just a desirable property for a localisation system, since it helps in the reduction of the system consumption, in the minimization of external infrastructures and in the relaxation of the assumptions to be made on the environment. In line with the considerations above, localisation throughout this thesis is not seen as the process that always has to maximise the accuracy of the estimated robot pose. On the contrary, localisation is considered as the process that minimises an objective function related to the infrastructure’s cost, to the power consumption and to the computation time, being subject to some requirements on the localization accuracy.