Training character controllers from motion capture data using Motion VAEs

What is it about?

We combine reinforcement learning and learned motion model to generate high quality kinematic character animations. We use reinforcement learning to train control policies for various locomotion tasks on top of a single motion model.

Featured Image

Photo by Cristiano Pinto on Unsplash

Photo by Cristiano Pinto on Unsplash

Why is it important?

We show that the two-step process of model and control is a more flexible alternative to direct-prediction methods in kinematic animation. We can train better control policies by answering "what the character can do?" and "how the task should be solved?" separately.