Knowledge-Powered Inference of Crowd Behaviors in Semantically Rich Environments

Abstract

Interactive authoring of collaborative, context-dependent virtual agent behaviors can be challenging. Current approaches often rely heavily on users’ input, leading to cumbersome behavior authoring experiences and biased results, which do not reflect realistic space-people interactions in virtual settings. To address these issues, we generate an ontology graph from commonsense knowledge corpus and use it to automatically infer behavior distributions that determine agents’ context-dependent interactions with the built environment. By means of a natural-language interface, users can interactively refine a building’s design by adding semantic labels to spaces and populating rooms with equipment following suggestions that the system provides based on commonsense knowledge. Based on the chosen setup, an authoring system automatically populates the environment and allocates agents to specific behaviors while satisfying a behavior distribution inferred from the ontology graph. This approach holds promise to help architects, engineers, and game designers interactively author plausible agent behaviors that reveal the mutual interactions between people and the spaces they inhabit.