In this paper, an unsupervised deep learning-based framework based on dual-path model-driven auto-encoders (AE) is proposed for angle-of-arrivals (AoAs) estimation in massive MIMO systems.