Planning and adaptation of virtual human character motion

Abstract

This thesis covers some work on planning and adapting motions for a virtual 3d human character. The motivation is to use the physical features of the motion to drive both the planning of motion sequences and the subsequent editing of the motions to fit with necessary adaptations to the placement of contacts between the character and the geometric environment. We present a scheme for planning sequences of motion in real time based on unrolling a graph of motions and testing availability of support for contacts. The contact model uses raycasts into the geometric scene to check for locations that approximately conform to predefined hand grip and foot contacts. Short-horizon planning (on the order of a few seconds of animation) can be done in real-time given a static geometry pre-process, though with a high memory cost. Since the foot and hand placement for these motions is allowed to vary somewhat, we look at physically-based adaptation of animations to accommodate altered contact locations and mass-centre positioning. In the context of a vertex-based character representation, with constraints defining the character skeleton at each animation frame, we introduce some novel constraints to describe dynamic quantities by finite differences over several frames. The method gives visually acceptable results for smaller adaptations to external constraints, resulting in more realistic motion than with inverse kinematics (IK) constraints and acceleration smoothing alone. However, the solver is not guaranteed to converge and tends to become unstable for larger deformations. We also explore an existing method for data-driven IK using principal geodesics analysis of an input set of motion capture data. We give full implementation details, and introduce an idea to improve the performance of the technique as a general full-body IK solver. The novel idea is to split the character skeleton into regions, extract the principal geodesics for each region and then recombine them in the IK solution. This gives more stable poses and allows a greater range of novel poses to be produced from the same input data set.