Autonomous, Explorative Motion Generation

Keyframe Animations of Virtual Humans via Motion Data Learning

Abstract

Traditional keyframe techniques for humanoid animations require skillful operations in designing postures and motion curves. On the other hand, recent methods introduce motion capture data for creating natural behaviors, and their reuse becomes an important technical issue. This paper extends our previous learning methodology for keyframe animations by reusing motion data. Plausible postures at all keyframes are extensively searched by using data-centric objective function, and they are interpolated with motion curves estimated from both theoretical model and referential clips of actual human movements. This new technique is well suited to efficiently generate complicated motions of a whole body.

Publications

1. Tomohiko Mukai, Shigeru Kuriyama, and Toyohisa Kaneko, "Motion Generation of Virtual Human with Hierarchical Reinforcement Learning", Electronics and Communications in Japan (Translation from Trans. IEICE), Part 3, Vol.87, No.11, pp.34-43, 2004.
2. Tomohiko Mukai, Shigeru Kuriyama, and Toyohisa Kaneko, "Motion Generation via Hierarchical Feature Learning of Motion Data", Computer Graphics and Imaging, pp.31-36, 2003.
3. Tomohiko Mukai, Shigeru Kuriyama, and Toyohisa Kaneko, "Natural Human Animation via Learning with Dynamic Manipulability", Computer Graphics International, pp.272-275, 2003.
4. Tomohiko Mukai, Shigeru Kuriyama, and Toyohisa Kaneko, "Extensive and Efficient Search of Human Movements with Hierarchical Reinforcement Learning", Computer Animation, pp.103--107, 2002.

Result 1: Spearing motions

-Original motion

Captured motion (MPEG-1: 623KB)

-Same constrained locations as motion data

-Change of the constrained locations

Reward function from motion data + Keyframe interpolation using acceleration template
(MPEG-1: 623KB)

Result 2: Clambering motions

-Original motion

Captured motion of climbing a ladder(MPEG-1: 1416KB)

-Higher constrained locations

-Change of constrained sequece of upper limbs

Reward function from motion data + Keyframe interpolation using acceleration template
(MPEG-1: 989KB)