Modeling rhythmic physical exercises

In the framework of our ongoing project [2, 3, 4], we aim at developing a Reactive Virtual Trainer (RVT), an intelligent virtual agents (IVA) who can not only perform exercises, but adjust those, and interweave the physical performance with explanations, correcting remarks and social chat. For an IVA in such a role where the demonstrated motion is to be copied by the user, the subtle, physically correct performance of the exercises she is to show is of major importance. The RVT should be flexible in two aspects: what kind of exercises are to be done, put together form a repertoire of simple motions; and how the exercise is to be performed, concerning tempo and intensity. Because of this envisioned flexibility and adaptivity, the RVT cannot rely directly on a mocap repertoire, but should have a model-based framework for both structural ‘design’ of the exercises and its performance details. In this paper we will show how we model these two aspects, based on analyzing real-life corpora of physical exercises.