Experiential robot learning with deep neural networks

摘要

Children exhibit learning behaviors such as repeatedly conversing with and asking their parents, or adults, about intriguing things. Our aim is to enable this behavior in robots, particularly Curiosity. We lay ground to a methodology that enables and facilitates that kind of open-ended learning. In this paper, we propose a method called Experiential Robot Learning to develop open-ended, developmental robotics by leveraging Deep Learning techniques. Our work aims at creating agents capable of continuous self-improvement through autonomous learning, using neural networks. Artificial Neural Networks have been around for decades. In the past few years, however, they attracted substantial attention by outperforming traditional Machine Learning approaches in accuracy and flexibility. We investigate the concept of interactive coaching, where a robot can ask "what is this called" when a previously unknown object is encountered. Secondly, we investigate the capability of a robot to move around a new object and view it from several perspectives. A physical humanoid platform, NAO robot, is used to evaluate our method. The robot communicates in natural language with a human coach to identify and label known and unknown objects. In the first experiment set, a human presents 3 objects to NAO's camera. In the second set, NAO autonomously walks around 13 objects to see them from all angles. Results show that the robot autonomously expands its own knowledgebase by learning to recognize new objects and generalize learnings, through human coaching.