The action models that are learned from studying human performances typically represent some form of prototypical performance, and also characterize the ways that the human’s performance of the action tend to vary over multiple performances due to external influences or stochastic variation. The original motivation for the work is for applications involving skill trans-fer from humans to robots by human demonstration. Instead of explicitly programming a robot to perform a given task, the object is to learn from example human performances a model of the human action skill which a robot may use to perform like its teacher. Action models are also useful for applications in which skills are transferred from a human expert to a student. A distance metric from an appropriate action model learned from human performances could be used to build an improved criterion for the naturalness of the simulated motions.