Towards Fluid Human-AI Collaboration

Abstract

In this presentation, we explore the concept of fluid collaboration (FC) which describes highly flexible, real-time teamwork that relies on dynamic task and role allocation, as commonly observed in everyday human collaboration. We argue that AI agents need to leverage both online Theory-of-Mind inference and natural language capabilities to support the required mix of implicit and explicit task coordination. We present the "Cooperative Cuisine" testbed-an extended Overcooked-like environment-allowing for exploring the design of cooperative agents and adaptive human-AI teams to achieve FC.