Leveraging Generative AI to Create Lightweight Simulations for Far-Future Autonomous Teammates

Abstract

As the domain of AI advances, the design and capability of human-AI teams are becoming increasingly complex. Unfortunately, this complexity has increased the pace at which research needs to be performed. On the one hand, low-fidelity survey-based experiments have provided an opportunity for rapid human-AI teaming research. High-fidelity research studies that use full-fledged simulations remain relevant, but their development overhead often slows the pace of research. This article proposes a system design that splits the difference to explore human-AI teams at a medium fidelity that allows for rapid prototyping from researchers and interaction from participants. The proposed platform consists of a predictive simulation engine that uses generative AI to ingest, modify, and predict simulation states. Researchers can describe teammate capabilities, environments, and goals, which can be stored in a traditional JSON game state. The proposed simulation provides an interactive opportunity to explore modern and far-future HATs.