A group of researchers from the University of Edinburgh, together with Method Studios, have crafted a virtual locomotion engine using neural networks. The result is some highly accurate motion for virtual characters across tricky terrain, and without the astronomical sets of static animations.
The process uses machine learning to take set, cyclical animations, match them up against different terrain heightmaps, and gauge its own success. The project is focused on locomotion — hence the cyclical nature of the animations, which when the system is trained, can stop the limb's cycle when it hits the ground. Non-cyclical animations could make the system usable for other things, such as attacks.
The abstract goes as follows:
The entire network is trained in an end-to-end fashion on a large dataset composed of locomotion such as walking, running, jumping, and climbing movements fitted into virtual environments. Our system can therefore automatically produce motions where the character adapts to different geometric environments such as walking and running over rough terrain, climbing over large rocks, jumping over obstacles, and crouching under low ceilings.
Apparently after the system is trained, it only takes a few milliseconds of execution time and a few megabytes of memory, even if it took gigabytes of motion data to train it. That's amazing, even if it might be initially limited to world-traversing before it's expanded to deal with different kinds of situations.
You can read more about the system in this paper.
Imagine how much Geralt and Roach would benefit...