Autonomous robot navigation is typically considered a geometric problem, in which the robot must identify the geometry of the environment to plan a collision-free trajectory. However, a purely geometric view of the environment may be insufficient for many problems in robotic navigation. For example, geometry-based navigation methods may cause a robot to avoid an area blocked by plant leaves, believing it to be impassable. So it is a fact that going beyond geometric approaches to achieve greater potential is essential. Thus, making use of methods that enable self-supervised learning about the navigability of rugged environments from the robot's previous experiences and sensory data provides a prediction of traversability. However, deep learning models that make this prediction can be investigated and improved through ablative studies, being differential in network causality and the main focus of this project.
News published in Agência FAPESP Newsletter about the scholarship: