Learning force fields for limb control in character animation

Abstract

This work applies vector-field-based limb control to physically simulated character animation. Vector fields of joint torques, defined over joint angle space, have been shown in the neuroscience literature to be able to generate simple reaching motions in two dimensions when linearly combined with time-varying weights. In this work, three-dimensional versions of such "basis" fields generate a similar diversity of realistic motions, while maintaining tractable control through a concise parameter space. A decomposition algorithm is presented that takes a set of limb motions and generates a set of torque fields, whose various linear combinations can generate each member of the input set, and other plausible motions. Unlike other motion extrapolators, the system extrapolates torque control programs instead of the motion paths themselves, retaining Newtonian physical correctness.