Deep Learning of Neuromuscular Control for Biomechanical Human Animation

Abstract

Increasingly complex physics-based models enhance the realism of character animation in computer graphics, but they pose difficult motor control challenges. This is especially the case when controlling a biomechanically simulated virtual human with an anatomically realistic structure that is actuated in a natural manner by a multitude of contractile muscles. Graphics researchers have pursued machine learning approaches to neuromuscular control, but traditional neural network learning methods suffer limitations when applied to complex biomechanical models and their associated high-dimensional training datasets. We demonstrate that “deep learning” is a useful approach to training neuromuscular controllers for biomechanical character animation. In particular, we propose a deep neural network architecture that can effectively and efficiently control (online) a dynamic musculoskeletal model of the human neck-head-face complex after having learned (offline) a high-dimensional map relating head orientation changes to neck muscle activations. To our knowledge, this is the first application of deep learning to biomechanical human animation with a muscle-driven model.