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

UCLA

UCLA Electronic Theses and Dissertations bannerUCLA

Core Training: Learning Deep Neuromuscular Control of the Torso for Anthropomimetic Animation

Abstract

Despite its importance, the core of the human body has to date

received inadequate attention in the computer graphics literature. We

tackle the challenge of biomechanically simulating and controlling the

torso, including of course the spine, in its full musculoskeletal

complexity, thus providing a whole-body biomechanical human model with

a full set of articular degrees of freedom actuated by many hundreds

of muscles embedded in a finite-element soft tissue simulation.

Performing skillful (non-locomotive) motor tasks while bipedally

balancing upright in gravity has never before been attempted with a

musculoskeletal model of such realism and complexity. Our approach to

tackling the challenge is machine learning, specifically deep

learning. The neuromuscular motor control system of our virtual human

comprises 12 trained deep neural networks (DNNs), including a core

voluntary/reflex DNN pair devoted to innervating the 443 muscles of

the torso. By synthesizing its own training data offline, our virtual

human automatically learns efficient, online, active control of the

core musculoskeletal complex as well as its proper coordination with

the five extremities---the cervicocephalic, arm, and leg

musculoskeletal complexes---in order to perform nontrivial motor tasks

such as sitting and standing, doing calisthenics, stepping, and golf

putting. Moreover, we equip our virtual human with a full sensorimotor

control system, thus making it autonomous. Afforded suitable NN-based

machine perception, our model can also visually analyze drawings and

manually sketch similar drawings as it balances in an upright stance

before a large touchscreen display.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View