Subtle social maneuvering for MetaHumans

Abstract

Recent advances in computer graphics have made it possible for virtual environments to feature realistic lighting and 3D models. Such 3D models can be achieved using the MetaHuman framework, which allows for creating photorealistic virtual characters. With these advances the focus starts to shift for better character locomotion and AI behavior. The characters must be able to showcase behavior that are closer to that of a human being. If the characters does not exhibit human like behavior than a user observing the character might feel at unease with the character, a phenomenon known as the uncanny valley. This thesis presents a system for virtual agents capable of performing realistic non-verbal social maneuvers in a virtual environment. Built on the MetaHuman framework in the Unreal Engine, the system integrates high fidelity character models with independently controlled body regions including gaze, head, chest, and a novel feet orientation subsystem allowing agents to express nuanced social behaviors such as low- and high-commitment interactions. A behavior component coordinates these body motions according to social dynamics, while a group conversation system manages turn-taking and multi-agent coordination within conversation circles. The design is grounded in established social theories by Scheflen and Kendon, particularly concerning interpersonal spacing, body regions, and F-formations. Evaluations demonstrate that the system not only supports believable agent behavior but is also flexible and easily configurable, enabling rapid adjustments to orientation speed, target selection, and social engagement parameters. This work contributes to the development of socially intelligent virtual agents for applications in games, virtual reality, and interactive simulations, bridging the gap between theoretical social behavior and practical, real-time implementation.