Learning Animation

 Creating realistic looking animations by hand is a difficult and very labour intensive task. Motion capture data can
help to ease this problem but to create continuously animated virtual characters remains a challenging task. In this
project machine learning methods such as reinforcement learning or expectation maximisation are explored in order
to extract information from motion capture data in order to “teach” virtual characters how they should move in
order to appear natural. Instead of defining keyframes and constraints we want to animate virtual character by
giving them more
abstract goals. Such goals can be defined quite easily for sports tasks as for example for a boxer it would be to hit a
boxing opponent character as often and hard as possible and to avoid as many hits as possible. However, it will be
more interesting and challenging to define goals for socially interacting virtual characters in order to enable them to
interact for example like humans would when they meet for example in a bar.
Reinforcement Learning http://www.cs.ualberta.ca/~sutton/book/the-book.html