Object Manipulation from Simplified Visual Cues

Humanoid robotics in general, and human–robot interaction in particular, is gaining new, extensive fields of application, as it gradually becomes pervasive in our daily life. One of the actions that humanoid robots must perform is the manipulation of things (reaching their arms for objects, grasping and moving them). However, in order to do this, a robot must first have acquired some knowledge about the target object and its position in space. This can be accomplished with a perceptual approach. The developed system described in this thesis is based on the CAMSHIFT visual tracker and on a 3D reconstruction technique, providing information about position and orientation of a generic, model-free object that moves in the field of view of a humanoid robot platform. An object is perceived in a simplified way, by approximating it with its best-fit enclosing ellipse. After having computed where an object is currently placed in front of it, the robotic platform can perform reaching tasks. Experiments obtained with the robot arm of the adopted platform are discussed.