We present a mobile robot that autonomously generates behaviors to calibrate its intuitive-physics engine, also known as the Game Engine in the Head (GEITH). Most POMDP and Active Inference learning techniques operate in a closed world in which the set of states is defined a priori. However, implementing an innate GEITH and a set of interactive behaviors allowed us to avoid these limitations and design a mechanism for information search and learning in an open world. The results show that over a few tens of interaction cycles, the robot's prediction errors decrease, which shows an improvement in the GEITH calibration. Moreover, the robot generates behaviors that human observers describe as playful.