Scaling Empowered Agents

<p>This paper introduces a scalable algorithm called <strong>Skills Estimate Empowerment (SEE)</strong> that bridges unsupervised skill learning and empowerment maximization to train intelligent agents in complex, open-ended environments. By treating mutual information skill learning as a lower bound for potential empowerment, the method utilizes a <strong>two-stage policy</strong> where a potential policy navigates toward controllable states while effective skill policies explore local options. Through experiments in challenging settings such as <strong>Craftax-Classic</strong>, <strong>AntSoccer</strong>, and <strong>BuilderBench</strong>, the authors demonstrate that this approach successfully overcomes previous computational limitations. Consequently, the trained agents exhibit <strong>emergent behaviors</strong> like resource gathering, tool crafting, environmental modification, and survival without relying on external rewards.</p>