Beyond saccadic movements for interactive perception in robotics

Abstract

This thesis takes inspiration from biology to replicate in robots the oculomotor strategies humans and other species use to acquire visual information. Firstly, it establishes the foundations of the work by reviewing current knowledge on the primate oculomotor system, the state of the art on adaptive learning in artificial neural networks, and the main types of robotic heads that would enable imitation of eye movements. Afterwards, the thesis goes in detail on selecting a specific head configuration, developing algorithms for the generation of accurate saccades, and developing internal models that are generalizable across different robotic instantiations. The acquired knowledge is then applied to the specific task of using saccadic gaze shifts to extract depth information from the scene.