We introduce a human-aware affective policy generation framework aimed at enhancing long-term interactions between humans and robots. The framework models emotional dynamics through a multi-modal input system, including gaze, speech tone, and proximity cues. Policies are generated using affective computing principles and reinforcement learning to adapt interaction strategies in real-time. Empirical evaluations demonstrate increased user engagement, social trust, and interaction consistency over time. The proposed system contributes to building socially intelligent robots capable of forming sustainable relationships in caregiving and service contexts.