A robot system needs to autonomously acquire new knowledge through interaction with the environment. The knowledge can be acquired only if suitable perception-action capabilities are present: a robotic system has to be able to detect, attend to and manipulate objects in the environment as well as interact with people and other robots. We present our work in the area of vision based sensing and control with specific objectives on attention, segmentation, multisensory control and learning with a focus on what and how to represent to achieve intelligent behaviour in robots.