From self-driving cars to vision-based robotic manipulation, emerging technologies are characterized by high-dimensional measurements of complex physical systems. Oftentimes, these systems lack simple models suitable for control design. Moreover, the high-dimensional measurements typically capture irrelevant dynamics that have nothing to do with the control objective (e.g. background movements in a video). This motivates my research into efficient reinforcement learning algorithms that can control complex systems with high-dimensional observations (e.g. images), while being robust against irrelevant dynamics.