Skip to main content
eScholarship
Open Access Publications from the University of California

Deep Learning of Neuromuscular and Sensorimotor Control with Biomimetic Perception for Realistic Biomechanical Human Animation

  • Author(s): Nakada, Masaki
  • Advisor(s): Terzopoulos, Demetri
  • et al.
Abstract

We introduce a biomimetic simulation framework for investigating human perception and sensorimotor control. Our framework is unique in that it features a biomechanically simulated musculoskeletal human model actuated by 823 muscles. The anthropomorphic model has two human-like eyes whose retinas contain spatially nonuniform arrangements of photoreceptors. The sensorimotor control system of our human model comprises a set of 15 automatically-trained, fully-connected deep neural networks. Two networks control the saccadic eye movement functionality of its binocular, foveated perception system. The remaining networks achieve neuromuscular control of the skeletal muscles. One network controls the 216 neck muscles that actuate the neck-head biomechanical complex, producing controlled head movements. In our prototype model, 3 networks control each limb; in particular, the 29 muscles in each of the two arms and the 39 muscles in each of the two legs. Thus, the virtual human demonstrates eective sensorimotor control of its eyes, head, and four limbs driven exclusively by visual perception to achieve a nontrivial motor task. We also demonstrate that its foveated perceptual system is capable of appearance-based recognition.

Main Content
Current View