Smart Objects for Attentive Agents

Christopher Peters, Simon Dobbyn, Brian MacNamee, Carol O'SullivanWSCG 2003

Introduction

Creating believable autonomous agents in virtual environments requires balancing high-level planning with nuanced low-level animation. Traditional approaches often result in "robotic" behavior, where agents lack realistic gaze or fail to coordinate naturally with others. This paper presents an enhanced "Smart Object" framework that addresses these issues by embedding interaction logic and attention metadata directly into virtual objects. By offloading social coordination and gaze behavior from the agents to the objects themselves, the system enables synchronized multi-agent animations with reduced computational overhead.

The Smart Object Framework

The authors extend the traditional Smart Object paradigm by introducing spatial and logical constraints that guide agent behavior. Central to this are User Slots, which define precise locations and orientations for agents, preventing spatial conflicts during multi-agent interactions.

To facilitate the creation of these objects, a bespoke authoring tool was developed as a 3D Studio Max plugin. This tool allows artists to visually place user slots and "Attention Points"—specific regions of a mesh that should draw an agent's gaze—and export this data as XML for use in simulation engines.

The custom 3D Studio Max plugin allows designers to visually define user slots and attention points on a bar counter model.

Synchronized Social Interaction

A key innovation is the use of Usage Steps to manage social synchronization. Interactions are decomposed into state machines where steps can include "wait" conditions. These conditions are triggered by the progress of other agents in different slots of the same object. For example, in a bar scenario, the "Barman" agent waits for the "Customer" to complete an "Order" step before proceeding to "Pour." This object-centric coordination removes the need for a centralized "scene brain" to manage complex social exchanges.

State machine synchronization between a Barman and a Customer, managed by the smart object.

Gaze and Attention Management

To improve realism, the framework includes a dedicated Gaze Manager that handles three distinct gaze types:

  • Glance: Subtle eye movement with minimal head involvement.
  • Look: Coordinated eye and head movement.
  • Stare: Full orientation of the head and spine toward a target.

By tagging object faces with "Attention Points," designers can ensure agents look at relevant parts of an object (e.g., a label or a handle) at the appropriate time. This manifestation of attention makes agents appear "aware" of their surroundings, significantly increasing perceived intelligence.

Results and Future Work

The system was validated through a complex "Bar Scene" simulation. Qualitative results showed that the gaze system effectively differentiated between interaction types, making agents appear more human. The authoring workflow proved efficient, allowing designers to embed complex logic without deep programming knowledge.

Future work aims to address current limitations, such as handling "consumable" objects that change state or spawn new items (like a poured drink). Additionally, the authors suggest automating the placement of attention points using geometric or texture saliency analysis to further streamline the design process.

Summary updated

BibTeX
@inproceedings{peters2003smart,
  author = {Christopher Peters and Simon Dobbyn and Brian MacNamee and Carol O'Sullivan},
  title = {{Smart Objects for Attentive Agents}},
  booktitle = {WSCG},
  year = {2003},
}
Figure