Only certain objects roll well, and to make them roll requires matching the robot's action to the object's pose in an suitable manner. The robot can locate familiar objects through vision and tap them from different directions using kinematic control. In order to get the rolling distance, iCub segments the object and finds its orientation to then track it using a particle filter. To learn the rolling affordance the Least-squares support vector machine (LSSVM) technique is employed. Once iCub has explored this affordance it is asked to tap the object to attain the maximum rolling distance.