Abstract
Dynamic response is a technique for employing a physical reaction to
an animated character. The technique utilizes a database of reactions
as example motions to transition to following a dynamic simulation of
an interaction. The search for the example to foll ow has been the
stumbling block for bringing such a system into real time applications
and in this paper, we address that issue by proposing a number of
speed-ups which make the approach faster than real-time and appropriate
for an electronic game implementation. We accomplish this speed-up by
using a supervised learning routine which trains offline on a large set
of dynamic response examples and predicts online among the choices
found in the database. Also, we propose a near-optimal routine which
finds the alignment of the selected motion for the given scenario based
on a sparse sampling with an additional speed-up over the original
algorthim. With both of these changes in place, we enjoy a
tremendous speed-up with inperceptable difference in the final motion
compared to previous published results. Finally we offer a few
additional alternatives that allow the user to choose between quality
and speed based on their individual needs.
Publication
Riverside
Graphics
Lab
(link)
Anticpation
from Example
Zordan, V.B., Macchietto, A., Medina,
J., Soriano, M., Wu, C.C., Metoyer, R., Rose, R.
ACM Virtual Reality Software and Technology (VRST) 2007 - to appear.
[Paper]:
- pdf
(draft) [355KB]
[Videos]:
- mov1
[21.3MB]
