Robot learning, a research area at the intersection of machine learning and robotics, is driven by the goal of building robotic systems that ca learn from experience to perform complex tasks. Research in this area is motivated by challenges that arise from acting in stochastic unstructured environments, while receiving noisy and high-dimensional
input data such as images. Machine learning provides powerful tools for solving problems in robotics that, due to their complexity, cannot be addressed solely by engineering. For efficiency and due to the high cost of collecting data in robotics, most practical methods are model-based. In other terms, the robot explicitly learns geometric and mechanical models that can help identify objects and predict how they would move. In this talk, I will present my ongoing work in the field of robot learning, which is centered around three main axes: (1) learning physics models of objects, (2) learning visual and geometric models of objects, and (3), learning to manipulate objects from human demonstrations.