Learning Behaviors from Human Teachers by Generalizing Task-Relevant Features

This paper proposes a general method of robotic imitation learning. In this method, robots learn inner common features of demonstrations, which are largely different from each other, by analyzing the similarities among the features of the demonstrations. Adaptive generation methods are related to each feature. At the generation stage, given new task-relevant constraints, robots can generate motion trajectories, which still have the common feature learned from the demonstrations, to achieve the task-goals. This methodology is an opened framework which enables researchers to design features and feature related generation methods according to the application requirements. Three experiments are designed for robots to learn behaviors from human teachers, and the demonstrations given at the teaching stage are largely different from each other. Experimental results are given in this paper to verify the effectiveness of our proposed methodology.