Creating physically embodied agents using realistic human motion models

Abstract

Many simulations are populated with physically embodied agents capable of taking physical actions in the virtual world. Creating these agents, or virtual humans, is demanding; not only must the agents demonstrate visual verisimilitude, but they must plan and act in a way that is consistent with that of humans, especially for training simulations in which the participants are attempting to learn real-world skills. This article discusses an approach for adapting agent decision-making techniques to accurately model the physical capabilities of human subjects. To achieve this, the authors rely on human movement data acquired with a motion capture apparatus to build physically realistic models of human movement. To aid agents' planning, the authors construct a physical capability model for the agents, an accurate estimate of the time required for a real human to perform various movement sequences. A cost map over the space of agent actions is calculated by creating and stochastically sampling motion graphs assembled from the human data exemplars. The agents can use this cost model during the planning process to select between equivalent goal-achieving plans. This technique leverages highly accurate movement information acquired from human subjects to create agents that plan in physically realistic ways.