Incorporating active vision into the body schema

Many humanoid robots are built with active vision systems that allow them to foveate objects in an anthropomorphic manner that approximates human eye movement. Unfortunately, the act of foveating an object causes a change in the system's epipolar geometry, the projective relationship between the two cameras. Interestingly, however, one solution to this problem has a strong relationship with another important problem in humanoid robotics, that of learning the body schema. In this work, we show the relationship between the two problems, building a model that incorporates the active vision system into the body schema. Such a model could be used to both facilitate stereo reconstruction in stereo active vision systems, and to build a relationship between what is seen in the visual field and the underlying kinematics of the robot. This will facilitate development of richer, more life-like, more communicative humanoid robots.