Context-Independent Task Knowledge for Neurosymbolic Reasoning in Cognitive Robotics

Abstract

One of the current main goals of artificial intelligence and robotics research is the creation of an artificial assistant which can have flexible, human like behavior, in order to accomplish everyday tasks. A lot of what is context-independent task knowledge to the human is what enables this flexibility at multiple levels of cognition. In this scope the author analyzes how to acquire, represent and disambiguate symbolic knowledge representing context-independent task knowledge, abstracted from multiple instances: this thesis elaborates the incurred problems, implementation constraints, current state-of-the-art practices and ultimately the solutions newly introduced in this scope. The author specifically discusses acquisition of context-independent task knowledge from large amounts of human-written texts and their reusability in the robotics domain; the acquisition of knowledge on human musculoskeletal dependencies constraining motion which allows a better higher level representation of observed trajectories; the means of verbalization of partial contextual and instruction knowledge, increasing interaction possibilities with the human as well as contextual adaptation. All the aforementioned points are supported by evaluation in heterogeneous setups, to bring a view on how to make optimal use of statistical & symbolic applications (i.e. neurosymbolic reasoning) in cognitive robotics. This work has been performed to enable context-adaptable artificial assistants, by bringing together knowledge on what is usually regarded as context-independent task knowledge.