ABSTRACT
Human behavior is a key factor in human-robot interactions. It is important for a robot to understand human behavior in order to 1) correctly predict the human motion in the future so as to safely and efficiently plan its own motion, and 2) learn skills to finish various tasks from human demonstration. In the first case, the robot only mentally simulates human behavior to make the prediction. In the second case, the robot needs to physically execute the learned skill models from human demonstration. The major limitation of existing imitation approaches for prediction is that the learned model is not individualized, which may not be able to track individual differences and time-varying behaviors. The major limitation of existing imitation approaches for skill learning is the lack of physical interpretation and stability proof, which is crucial for implementation on industrial robots. This chapter discusses methods to improve prediction and skill learning through imitation. The imitation for prediction through offline behavior classification and online model adaptation is discussed first, followed by the imitation for action (or skill learning) in a model-free framework.
