Learning and inference methods offer the opportunity to transform how we control and model complex dynamical systems, ranging from robotic platforms to power systems in a building. In this talk, I will first discuss my work in applying learning methods to robotic systems, namely a quadruped robot and an autonomous car. In both cases, I will show how learning approaches are able to obtain impressive performance on dynamic tasks.  Looking forward, I will highlight another domain where these data-driven approaches can have a profound impact: energy and sustainability.  In particular, I will look at the task of analyzing aggregate energy consumption in a building to discover and separate different appliances; I will discuss new probabilistic inference methods that enable far better performance than was previously possible.  I will also discuss new Reinforcement Learning approaches to control a small physical wind turbine system, and demonstrate that these allow us to generate 30% more power than purely model-based control.