PLANNING TOWARDS EFFECTIVE MOBILE ROBOT NAVIGATION IN UNSTRUCTURED ENVIRONMENTS

Abstract

This thesis presents our work on planning pipelines and frameworks towards effective mobile robot navigation in unstructured environments. Recognizing the current limitations of existing related works on autonomous navigation subject to issues such as environment data inadequacy, complex geometries, and cluttered obstacles, etc., we try to tackle the corresponding challenges from the planning perspective, including efficient information acquisition in unknown environments, practical path finding in cluttered environments, active path clearing in presence of movable obstacles, and the associated dynamic object pushing manipulation for reconfiguring local space with collision avoidance. Basically, we propose an optimized next view planning framework for exploration and reconstruction of unknown environments, an obstacle-guided informed path planning paradigm for navigation in cluttered environments with narrow passages, a space reconfiguration planning pipeline for active path clearing through movable obstacle manipulation, as well as an efficient learning-based approach for collision-aware, non-prehensile, planar pushing manipulation of box-shaped objects, using target-driven end-to-end model.