For decades, the most capable robotic systems have been a product of brilliant control engineering. We believe that taking robotics further, to its full potential in fields ranging from everyday assistance to datacenter maintenance to agriculture, requires machine learning in addition to control. In close collaboration with computer vision and reinforcement learning teams, our group does research at the intersection of large-scale multimodal representation learning, imitation learning, reinforcement learning, computer vision, and NLP in order to make training robots to perform new tasks natural and effortless. Among the physical meta-skills robots have yet to master, manipulating objects with human-level dexterity is a challenge — and a promise — bar none. To that end, it is the focus of our group’s main effort, Project Dexter.