Motion matching started a revolution in the way developers create runtime animation controllers for video game characters, freeing developers from the burden of manually crafted motion trees. Games like 'For Honor', 'UFC' and 'Last of Us' are showing the great benefits in terms of realism and animation quality this technique can provide. Still, motion matching can only choose poses from an animation database, with no ability to generate new ones. Machine learning can help you go a step further, from motion matching to actual motion synthesis.

This session will cover state-of-the-art techniques (Phase-Functioned Neural Networks, and Mode-Adaptive Neural Networks) that use neural networks to synthesize motion from examples, explicitly calling out important architecture and implementation details, and spark a discussion on how this technology can be used in a modern game development pipeline.

Takeaway

Attendees will learn about recent academic approaches to using neural networks for character animation, be provided insight about what works based on practical experience and be guided in avoiding common pitfalls when implementing such systems.

Intended Audience

This talk is aimed at programmers interested in exploring machine learning solutions to character animation problems.
A basic understanding of simple feed-forward neural networks is required, together with a solid background in linear algebra and 3D transformations.