In the wake of increasing interests in artificial intelligence, problems such as massive training data requirements and single task learning constraint are unavoidable. Thus we propose a Learning from Demonstration (LfD) system. The LfD uses deep learning techniques including image recognition, action and facial recognition, etc. As such, robots learn actions directly from human instead of training data.

To allow robots learn from human demonstration, action recognition and object detection are integrated in the LfD. An action is recognized and fragmented into sub-actions using I3D deep learning architecture. YOLOv3 is used to detect and localize the objects in the environment. An action base is therefore developed based on action recognition and object detection, such that the robot can learn demonstrated actions.

The proposed LfD can be employed in indoor environments, especially in circumstances like home services. Because of the capabilities that enable robots to learn from repeatedly human demonstrations, LfD can be employed in various different environments. Also, LfD combines actions with objects such that robots are able to learn the purposes of actions rather than only their appearances.