Selforganization of character behavior by mixing of learned movement primitives

The real-time synthesis of natural looking human movements is a challenging task in computer graphics and robotics. While dynamical systems provide the possibility to parameterize behavior in a flexible and adaptive way, the reconstruction of details of human movements with such systems is a challenge due to the large number of involved degrees of freedom. We present an approach for the synthesis of realistic human full-body movements in realtime that is based on the learning of motion primitives, or synergies, from motion capture data applying a novel blind source separation algorithm. By application of kernel methods we map such components onto low-dimensional dynamical systems that can be iterated in real-time, and which are integrated in a stable overall system architecture. We demonstrate how this model can be integrated with other key elements of computer animation systems, such as style morphing, synchronization with external rhythms or navigation. The performance of our approach was validated by self-organization of complex behaviors like dancing.