Automated Generation of Plausible Agent Object Interactions

John T. Balint and Jan M. Allbeck

Proceedings of Intelligent Virtual Agents 2015, Springer, page 295-309.

Abstract

To interact in a virtual environment, an agent must be aware of its available actions and the objects that can participate in these actions.Work such as Kallmann's SmartObjects allows for this information
to be encoded by a simulation author, but these encodings tend to be constrained to individual graphical models. Furthermore, there may not be agreement on the proper operational information between simulation authors. To create consistent object operational information, we have devised a method that uses natural language lexical databases to parse and assemble this information. The method uses a two step process: 1. The names of motion clips are disambiguated using their name and a list of keywords; 2. Objects are connected to the actions by resolving the operational elements of a given verb. This method is tested on several common, publicly available action sets and the accuracy and coverage of our method are measured.

Paper (pdf)