Active Learning for Real-Time Motion Controllers

Alex J. Champandard on August 3, 2007

Siggraph 2007 is around the corner, and that means new technology for game developers to play with! In terms of game AI, the most interesting part of the conference is character animation, as it promises to deliver realistic motion for lower investments.

One paper from University of Washington Animation Research Labs presents an active learning controller. There are two innovations here:

  • A kinematic character controller that can be built incrementally from motion-capture data.

  • An interactive workflow with the animator for easily adding new motion clips.

This combination makes it easier to build low-level animation behaviors, and visualize them immediately. It reduces turn-around times, and allows the animator to identify problems and correct them with new motion clips.

Screenshot of the movie

View or download the movie (MOV, 116 Mb).

Here’s the abstract:

This paper describes an approach to building real-time highly-controllable characters. A kinematic character controller is built on-the-fly during a capture session, and updated after each new motion clip is acquired. Active learning is used to identify which motion sequence the user should perform next, in order to improve the quality and responsiveness of the controller. Because motion clips are selected adaptively, we avoid the difficulty of manually determining which ones to capture, and can build complex controllers from scratch while significantly reducing the number of necessary motion samples.

Download the paper (PDF, 2.4 Mb):

Active Learning for Real-time Motion Controllers
Cooper, S. Hertzmann, A. Popovi?, Z.
ACM Transactions on Graphics 26(3)

Here’s a quick assessment of the technology based on how easy it would be to use for upcoming games.

Applicability to games: 6/10
Only AAA games that rely heavily on mo-cap would benefit. It requires a non-negligible budget for the equipment.
Usefulness for character AI: 7/10
Improves animation workflow significantly for human character animations. It provides better and more realistic mo-cap based behaviors.
Simplicity to implement: 1/10
The biggest hurdle is the motion capture equipment and interactive environment. But the technology itself also requires animation programming expertise.
Animator in front of a big screen.

Interactive Motion Capture Session

How do you think this kind of technology can be useful to game AI?

Discussion 1 Comments

gware on August 3rd, 2007

This framework is quite interesting for motion capture studios. It may be of a great help for "in-house" mocap studios, where you can spend a bit of time trying to refine animation controllers. I believe its a bit harder for studios which rely on mocap sessions and do not have the studio available for long. Usually there's not enough time for the mocap, thus studios may prefer asking for a full set of movements to be mocap'd and fine tune the raw data later. I'm not sure about the improvement on the workflow. Sure it helps to see what and how each mocap can be used, and being assisted during the mocap process. But, I beleive most animators and game companies knows more or less exactly what they need. Another point that's bothering me is that this does not take into account the person being mocap'd. I think that sports motion capture may be a bit tricky to do since you may want to do 8 hour of mocap straight but can't because of human factors. Sure you will be able to change the controller, and such. But where is the gain if you need additional manipulations each time you want to enter mocap data ? I like very much point 4.1 "selecting candidate tasks". This approach may pinpoint what's missing in the mocap set. For example, good transitions are usually a big concern thus the transition error metric can be really useful. This kind of software could really help animators with their mocap sets. Plus it does not require any special equipement (other than a good database, a bit of time to set the metrics, and a bit of computing power).

If you'd like to add a comment or question on this page, simply log-in to the site. You can create an account from the sign-up page if necessary... It takes less than a minute!