My research goal is to design robots which can develop the manipulation capabilities of human learners through exploration in natural environments. In particular, I am interested in manipulation of objects that the robot may have never seen before. To that end, I work on robot intuitive physics, a machine learning approach for efficiently modeling arbitrary world dynamics in terms of a space of latent physical quantities. Critically, and unlike previous approaches to parameter learning, my model representation adapts to the data observed by the robot. This provides an online bayesian perspective to system identification, and will (I hope!) pave the way to more sophisticated autonomous behavior for humanoid robots.
Lower-level robot stuff