Phase-Functioned Neural Networks for Character Control

For most games with animated 3D characters, a lot of time is spent linking the animations with the character’s motion, often using complex state machines that create webs of animation transitions.

This research, by Daniel Holden, Taku Komura, and Jun Saito, instead used a neural network to act as the character controller. The neural network is trained on a large dataset of animations and terrain data, taking gigabytes of data and combining it into a function that runs quickly and uses only a few megabytes of memory.

There’s been some past research in this area, but based on the video of their results their phase-functioned approach is very, very effective.

This is the exact kind of generative tool that can empower artists. It still needs the artistic input (that animation data has to come from somewhere) but it takes care of the very tedious work of combining all of those animations, freeing the artists to produce even more art. (And the technical artist can go improve some other tool.)

And, since the training is offline, rather than while the game is running, the risks of training a neural net can be supervised, so the game can ship with just the resulting locked-in function.

http://theorangeduck.com/page/phase-functioned-neural-networks-character-control