In human cognition, activity often plays a key role. Whenever there is ambiguity or uncertainty in processing stimuli, humans inevitably tend to take another look at an object, interact with objects, or simply ask "what?" when they don't understand what they are told. In my talk, I will focus on the various components that are required to implement active cognition on an autonomous robotic platform. In particular, such systems will need the reconstruction of stationary background, trajectories of moving foreground objects, as well as egomotion estimation and hand-eye coordination. But recognition of objects, motion patterns, events, and episodes plays an equally important role. Successful interaction with a dynamic scene demands to know about object and pose hypotheses to plan and perform useful next actions. Two enabling components will be presented in detail: Multibody Structure and Motion analysis (MSaM), and 3D object category models.

I will discuss several possible embodiments of such systems, including an active camera, an autonomous vehicle, and a humanoid robot, and present experimental results obtained in our lab. Besides a number of obvious applications and benefits, there is the hope for a better understanding of cognition through action by this research.