Supporting Character Realism (SCR)

Overview

Convincing virtual training environments and simulated worlds require rich populations of characters whose behavior feels consistent, socially aware, and grounded in the surrounding space. Traditional AI approaches either lavish heavy authoring effort on a single “main character” or populate crowds with simplistic, repetitive state machines that quickly break immersion.

In collaboration with Josh McCoy, Michael Mateas, and researchers at BAE Systems, we formalized the concept of Supporting Character Realism (SCR)—designing autonomous agents that work in tandem with human confederates and primary AI characters by exhibiting nuanced, contextually appropriate background social behaviors.

Architecture & Experimental Evaluation

• Reactive Planning in ABL: Authored modular behavior trees in ABL (A Behavior Language, a reactive planning language that compiles to Java), structuring behaviors into three reusable tiers:
  • Role-Unique Behaviors (Blue): Specialized social tasks specific to a character archetype.
  • Intra-Role Shared Behaviors (Green): Reusable sub-behaviors shared across tasks within a single role.
  • Cross-Role Universal Behaviors (Orange): Foundational navigation, gaze, proxemics, and conversational turn-taking shared by all agents.
• Distributed Simulation Bridge: Connected the Java ABL runtime to Second Life over a custom low-latency JSON socket protocol and C# sensor/actuator control console.
• Empirical Evaluation: Evaluated our proposed SCR realism metrics in a multi-patron virtual bar scenario (Three Avatars Walk Into a Bar) featuring Regular, Spring Breaker, and Waiter archetypes alongside human-controlled avatars, demonstrating that Supporting Character Realism is both empirically distinguishable and achievable with modular reactive planning.

Videos

Play

Screenshots & Figures