Coachable Agents for Interactive Gameplay (arxiv.org)

Recent research has introduced a groundbreaking framework for training AI agents, allowing for real-time coaching of stylish behaviors in interactive gameplay settings. Unlike traditional reinforcement learning, which typically results in a singular, optimized behavior through trial-and-error, this innovative approach incorporates universal value function approximators (UVFAs) alongside carefully designed training scenarios. The methodology was successfully implemented in popular AAA video games like Horizon Forbidden West and Gran Turismo, as well as in an open-source humanoid test domain. The results showcase that these agents can adapt to style requests while maintaining task performance across diverse environments. This development is significant for the AI/ML community as it offers users unprecedented control over agent behavior, enabling immediate adjustments based on preferences during gameplay. Such flexibility could revolutionize how players interact with AI in gaming, leading to more tailored experiences. The implications extend beyond gaming, suggesting potential applications in robotics and other AI systems where nuanced control over behavior is essential. Overall, the framework paves the way for more engaging and customizable interactions between users and AI agents, marking a notable advancement in the evolution of intelligent systems.