Motion generation of virtual human with hierarchical reinforcement learning

Abstract

AbstractIn the key frame method, which is a general method of generating the motion of a virtual human, a large number of postures must be specified for the virtual human in each key frame. A tremendous effort is required if this process is performed by a manual operation. This paper proposes a method of efficiently constructing the motion of the virtual human by specifying the position of the end effector in each key frame. In this method, the posture of the virtual human in each key frame is efficiently extracted from a tremendous number of candidate postures through hierarchical reinforcement learning. The virtual human model to be used is hierarchically organized according to the learning structure, thus partitioning the search space. Each component of the body retains a posture template with discrete joint angles. By the proposed method, composite continuous motion can be generated automatically by specifying the locus points of the fingertip or the toe in motion. In addition, the motion is generated with overall optimization based on an arbitrary condition. © 2004 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(11): 34–43, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20118