Recent work in robot learning has been largely driven by scaling model capacity, training data, and computation. Although scaling is important, my main research objective is to go beyond naive scaling. Once a capable robot policy is trained and deployed, it should not remain frozen. Instead, it should continue to learn from failures, accumulate experience, and improve its behavior over time. Humans learn in a similar way by continuously building internal models of how the world works. These world models allow us to predict the consequences of potential actions before executing them, reducing the amount of physical trial and error required to learn successful behaviors. Therefore, my research focuses on equipping robots to self-improve from experience, imagine possible futures, reason about their decisions, and continuously improve through both real and imagined experiences, where deployment becomes part of the learning process rather than its endpoint.