Robust Vision-Based Navigation in Extreme Environments
Inspired by the Hippocamal-Entorhinal 


Tomáš Musil


This thesis deals with bio-inspired monocular visual navigation and is divided into three main parts. First, we survey the research in cognitive neuroscience on human and animal navigation, compare it with the currently best robotic vision-based navigation systems, and identify core strengths of biological navigation which could be worth replicating in robotics. Second, we present a novel simulator that we purpose-build to simulate challenging scenarios and find the limits of existing vision-based systems, which often assume small-scale, static, and unambiguous environments. We also propose a method of evaluating navigation holistically, using the simulator. Third, we present a new vision-based autonomous navigation, map-building and exploration approach. We also show that by using large-scale spatial geometry instead of visual appearance, one can achieve robust multi-session localization even in a highly perceptually aliased environment. Finally, we demonstrate the exploration and safety-aware planning of the designed system both in simulation, and in the real world on an inexpensive unmanned aerial vehicle with limited sensing capabilities.

Multimedia material