Empowerment measures an agent’s ability to actively control its environment. This paper clarifies the relationship between skill learning and empowerment. While skill learning is not empowerment, we propose a simple, scalable recipe for empowerment maximization: empowerment corresponds to the number of executable skills from a given state, so maximizing it means reaching states that serve as good starting points for many skills. The recipe works with any underlying skill‑learning method. Prior work on empowerment largely examined tabular settings; our approach enables study at a scale where theoretically predicted but empirically elusive properties—such as resource gathering, environment modification, and survival—can be observed. Experiments show that empowerment maximization yields emergent behaviors, e.g., building a house in the game Crafter. The website and code are available at https://empowerment-geometry.github.io/scaling-empowerment/.
Review