Embodied Localisation and Mapping

Mobile autonomous robots have finally emerged from the confined spaces of structured and controlled indoor environments. To fulfil the promises of ubiquitous robotics in complex outdoor environments safe navigation is a key requirement. The research in the simultaneous localisation and mapping (SLAM) community has largely focused on using optical sensors to solve this problem. The fact that the robot is a physical entity which interacts with the environment has largely been ignored. In this thesis a method for localisation and mapping is proposed, which takes the embodiedness of robots into account. A Rao-Blackwellized Particle Filter is used with embodied measurement and prediction models to generate 3D map segments of outdoor environments. Using a hierarchical approach, these segments are combined in a pose graph with explicit global constraints. The errors of the optimized pose graph are back-propagated onto the local map segments, resulting in maps with smooth transitions along the path of the robot. The proposed method is experimentally verified in simulation and on two different physical outdoor systems. The results show that the approach is viable and that the rich modeling of the robot interaction with its environment provides a new modality for improving existing solutions and extending to scenarios where visual processing is not feasible