Experimental Studies in Pursuit of Experiential Robot Learning

Robots are currently not mature enough to be used in unconstrained environments (i.e. in the wild) because they cannot learn and thus cannot respond to new situations. Our hypothesis therefore is that the development of a methodology that permits experiential learning could allow robots to learn and therefore to succeed in novel situations. We developed a method called Experiential Robot Learning (ERL) that outlines how robots should be developed. Neural Networks (NN) provide a promising path towards ERL and this dissertation evaluates this promise. Experimental studies illuminated a problem with using NN for ERL: the need for a differentiable loss functions and architectures can’t always be satisfied, and the exploitative-nature of gradient-descent is not suitable to solve problems that require exploration. To address these shortcomings, we developed Local Search, a NN training approach that provides good results in the absence of a differentiable loss functions, or a loss function entirely, and on problems that require exploration. Our work paves the way for more advanced robot implementations adhering to ERL method.