PERCEPTION AND NAVIGATION OF LEGGED ROBOTS IN CHALLENGING ENVIRONMENTS

Abstract

In recent years, legged robots have gained popularity for navigating challenging terrains inaccessible to wheeled counterparts. However, they encounter difficulties in traversing various terrains, including obstacles like tall grass or low bushes that may seem impassable based on geometry but are navigable considering semantic properties. Conversely, geometrically passable obstacles like water puddles and sand can pose risks, while soft terrains like swamps may be hazardous despite geometric feasibility. Addressing these challenges, this thesis aims to enhance terrain sensing and motion planning for legged robots, improving navigation mobility, safety, and efficiency. This thesis introduced a vision-based traversability analysis, utilizing geometric and semantic features for a probabilistic traversability map. Two terrain collapsibility analysis approaches are proposed, employing tactile sensing and a digital twin approach. Navigation and locomotion control strategies are introduced to enhance safety and feasibility in path and motion generation. Successful experimentation in simulation and real-world tests validates the proposed methods, offering solutions to unsolved challenges in autonomous robotic navigation.