Lavatar
Learning Avatars in Virtual Worlds
Lavatar was a research project focused on building autonomous learning agents that operate as avatars inside virtual world environments, investigating how machine learning could enable avatars to acquire and adapt behaviors through interaction with rich, socially structured digital spaces.
Overview
As virtual worlds such as Second Life and online game environments grew rapidly in the late 2000s, they presented a unique opportunity for AI research: immersive, rule-governed spaces populated by human users and autonomous agents alike, where behavior could be observed, measured, and learned at scale. The Lavatar project (“Learning Avatars”) treated these virtual worlds not merely as entertainment platforms but as cognitive training grounds — environments where learning algorithms could be tested in social, physical, and goal-directed scenarios that mirror the complexity of the real world.
The central challenge Lavatar addressed was how an avatar agent, starting with minimal prior knowledge, could observe its environment, interact with human users and other agents, and autonomously acquire and refine behavioral competencies through experience. This connects the project to broader questions in cognitive science about how behavior and knowledge co-develop through environmental interaction — a theme central to the SNU Biointelligence Lab’s long-running program on machine learning from lived experience.
Technical Approach
Lavatar drew on the lab’s expertise in probabilistic and evolutionary machine learning:
- Bayesian behavior modeling: probabilistic inference over action sequences to represent and update beliefs about effective avatar behaviors in different in-world contexts
- Hypernetwork architectures: weighted random hypergraph structures encoding higher-order relations between environmental states, social cues, and behavioral outcomes — consistent with the lab’s broader hypernetwork research thread running through the MMG and SKT Hypernetwork projects of the same era
- Reinforcement learning: goal-directed avatar behavior acquisition through reward signals derived from in-world task completion and social feedback from human users
- Social interaction modeling: representing and responding to the behaviors of other agents (human and artificial) in shared virtual spaces, capturing the fundamentally social character of virtual world environments
A key motivation was that virtual worlds provide dense, continuous feedback — every action produces observable consequences, every interaction leaves a trace — making them ideal for evaluating incremental, online learning algorithms that must adapt without resetting to a blank slate.
Research Team
Principal Investigator: Prof. Byoung-Tak Zhang (Seoul National University, Biointelligence Lab)
Connection to Broader Lab Research
Lavatar sits at a productive intersection of several concurrent Biointelligence Lab projects:
| Related Project | Shared Theme |
|---|---|
| MMG (2007–2010) | Learning from naturalistic interactive environments |
| SKT Hypernetwork (2007–2008) | Hypernetwork architectures for behavioral learning |
| E-Learn (2009–2010) | Adaptive learning from ongoing user interaction |
| mLife (2010–2015) | Modeling behavior from continuous experiential data |
The Lavatar project extended the lab’s Bayesian and hypernetwork frameworks — developed for language and biological data — into the domain of embodied agent behavior in social virtual environments, anticipating the lab’s later work on cognitive agents and embodied AI.
Historical Context
The project was conceived at a moment when virtual worlds held significant research interest across AI, HCI, and the social sciences. Second Life peaked at over a million active users in 2009, and academic institutions worldwide maintained in-world presences for education and research. Lavatar positioned the SNU Biointelligence Lab within this wave, exploring how the controlled-yet-rich nature of virtual environments could accelerate progress on fundamental questions about autonomous behavioral learning — questions that would later re-emerge in the context of game AI, simulation-based robot training, and embodied AI research.