A human partner returns a specific response after a robot acts a specific social cue. We define this as interaction rules. The partner and robot continuously search for and co-create interaction rules as inspired by social games played between an infant and a caregiver. We propose a scheme composed of “making response prediction,” “confirming response prediction,” and “habituation/dishabituation to response prediction,” and developed a robot model composed of response-predictability and response-habituation. The robot generates actions, observes the partner's response, and get to predict them. It identifies relationships between its actions and the responses, and generates actions to confirm specific responses from the partner. The interaction is reciprocated as a result. After it is habituated to the responses, it inhibits the confirmation and generates other actions. This makes a chance for other rules. We conducted experiments in human-robot interaction using a ball based on this model to investigate whether response-habituation is needed or not, while response-prediction is required by the definition of interaction rules. As a result, analysis with causality measure proved that an appropriate response-habituation supports interaction reciprocation. Various patterns of interaction emerged, such as passing the ball back and forth, rolling and catching, feint passing, and role-reversal feint passing. Response-predictability increases when an interaction is reciprocated. Then response-habituation increases and the dyad quits the reciprocated interaction and search for another rule, indicating that the scheme and the model work.