Understanding mobile robots’ traversability over terrains is a crucial component of outdoor autonomous systems, since knowledge of their traversability helps robots to plan safe, fast, and energy-efficient paths. In deployments such as agriculture, forestry, or environment monitoring, a mobile robot can encounter terrains with a priori unknown traversability. The visual appearance and geometry of these terrains might be misleading, such as tall grass that appears as a rigid obstacle when only geometry is considered. The Thesis addresses these challenges by designing a self-improving traversability assessment system. The designed system follows the near-to-far paradigm, where the robot’s prior traversal experience is extended to untraversed terrains based on similarities in visual appearance and geometry. The Thesis is presented as a collection of four core publications that address three identified research challenges. The first challenge is focused on learning the traversal experience in a self-improving system, and represents a building block to solve the following challenges. The second challenge focuses on active traversability learning in mobile robot exploration, where the self-improving nature is realized by online decision-making concerning both where to learn the traversability and where to explore the spatial model. The third challenge extends the notion of traversability and thus the scope of the self-improving system through the description of the force to pass through the non-rigid obstacles.

Doctoral Thesis

Learning Traversability from Mobile Robot Experience

Miloš Prágr