A Comprehensive Study on Robot Learning from Demonstration

Abstract

The robots are set to penetrate into daily life of human beings. The purpose of a robot is to accomplish a goal within specified task constraints, thus reducing human effort as well as time. Since most of the end users are novice; the ability to use/adapt to a robot will vary. Besides that, since user references will be different, the task performance specifications/constraints would vary. An approach where an end user can teach a robot to perform desired tasks without the need of tedious reprogramming is learning from demonstration. This paper presents a survey of robot learning from demonstration. The problem formulation for LfD (Learning from Demonstration) is analyzed. The various modes of demonstration data acquisition and various data processing methods are discussed in an organized manner. Various learning approaches reported in literature are put forth. The issues of demonstration methods, suboptimal demonstrations, and evaluation metrics are highlighted. Lastly, the current issues which limit the adoption of LfD in real world scenarios are put forth in order to determine the future scope of work in this domain.