This paper introduces Generative Action Synthesis (GAS), a novel method for imbuing robots withhuman-like emotional expression during task execution. GAS leverages conditional WassersteinGANs (cWGANs) for action generation conditioned on emotional embeddings, guided by expertdemonstrations and refined via Hamiltonian Monte Carlo. A temporal-hierarchical transformer(THT) synthesizes actions while a Von Mises-Fisher mixture model (vMF-MM) resolves ambiguities.The framework also employs stochastic policy gradients for dynamic adjustment based on real-time feedback and task requirements, with Fokker-Planck-Kolmogorov equations ensuring emotionstability. This approach, integrating generative models with reinforcement learning and structuredemotional embedding, enables robots to exhibit a range of emotional behaviors, including anger,humor, and empathy, leading to more natural and adaptable human-robot interactions. Practicalimplications include advanced applications in caregiving, customer service, and other social domains,highlighting its significance in the development of emotionally intelligent robots