As socially assistive robots become increasingly integrated into everyday life, their ability to engage in emotionally intelligent interactions is essential for supporting long-term human-robot relationships. This work presents a reinforcement learning-based conversational model that dynamically adapts to the user’s emotional state across multi-turn dialogues. Using Proximal Policy Optimization (PPO), the system selects appropriate response types based on emotion labels extracted from user speech. A pre-trained emotion classifier (DistilRoBERTa) processes speech-to-text input, and the selected response type guides a GPT-based response generator. The model is deployed on Pepper, a humanoid robot, and evaluated through a two-week longitudinal user study comparing static and adaptive dialogue systems. Results demonstrate that the PPO-powered adaptive model significantly improves user engagement and reduces stress levels over time. This study highlights the importance of reinforcement learning for sustained emotional coherence and opens pathways for emotionally adaptive robots in mental health, education, and social support applications.