Neurosymbolic Robot Programming

Abstract Programming intelligent robots requires novel approaches that harness the learning abilities of deep neural architectures while allowing humans to specify, monitor and understand robot behavior. This dissertation conceptualizes robot programs as predictive models of robot tasks and proposes a neurosymbolic program representation that allows robots to learn and optimize their own behavior, while remaining parameterizable and interpretable by human users. It offers a blueprint for AI-enabled robot programming that bridges the apparent divides between symbolic and subsymbolic representations, designed and learned behavior, and capabilities and safety. The framework was evaluated on a range of robots and manipulation tasks from wire harness installation to kitchen clean-up and shopping assistance in retail stores.