Fast Robot Learning using Prospection and Experimental Knowledge : A Cognitive Approach with Narrative-Enabled Episodic Memories and Symbolic Knowledge

Humans employ data-efficient learning mechanisms to obtain new skills and to improve the existing ones. Robots have already replaced humans in terms of labor in performing repetitive and dangerous tasks in structured environments like factories. On the other hand, developing completely autonomous robotic systems for unstructured environments, like a household, is still a challenge for roboticists due to the infeasibility of programming every possible case and today's data-hungry machine learning approaches. In order to assist humans in such environments, I believe that robots should be able to gain and improve skills using human-like learning mechanisms regularly. For this purpose, I present a cognition-enabled fast learning framework in this dissertation which makes use of symbolic knowledge, episodic memories, and cloud robotics services along with a cutting-edge deep imitation learning methodology in order to reduce the dependency on big experiment data. Using this framework, robots can (1) imitate tasks demonstrated by a human demonstrator in virtual-reality, (2) adapt the actions of itself and others to new conditions, and (3) prospect which task parameters lead to the desired goal. To validate these abilities, I have provided some experimental results. These experiments were conducted with four different service robots in various kitchen environments. The human demonstrations were recorded inside a game-with-a-purpose using virtual-reality equipment. Such a setup enables roboticists to crowdsource their training data by eliminating the requirement of being in the same environment with the robot.