Hyperspectral remote sensing faces the persistent challenge of mixed pixels, where sub-pixel heterogeneity compromises analytical accuracy. Hyperspectral unmixing addresses this by decomposing mixed pixels into pure materials (endmembers) and their abundances, enabling precise sub-pixel information extraction. Deep learning, particularly autoencoder-based architectures, has emerged as a transformative approach, surpassing traditional physical models and shallow neural networks in modeling complex spectral mixing mechanisms. This survey systematically reviews autoencoder-based hyperspectral unmixing methods. First, the fundamentals of autoencoder networks are introduced, emphasizing their feature extraction, data reconstruction, and unsupervised learning capabilities, which align naturally with unmixing tasks. The evolution of model architectures is analyzed across three phases: 1) Single-network structures using basic autoencoders to validate feasibility; 2) Module-integrated networks incorporating convolutional, recurrent, or Transformer modules to enhance spectral-spatial feature modeling; and 3) Adaptive networks employing neural architecture search (NAS) or plug-and-play mechanisms for dynamic optimization.A critical focus is the integration of physical models with autoencoders. Early approaches imposed physical constraints (e.g., non-negativity, sum-to-one) on network inputs/outputs, while advanced methods now embed physical principles (e.g., radiative transfer models) into loss functions or network layers, achieving deeper synergy between data-driven learning and physics-based interpretability. Autoencoder-based unmixing has demonstrated remarkable progress, yet challenges persist in handling complex mixing scenarios, computational efficiency, and real-world adaptability. Future directions include: 1) Architectural innovation through fusion with transformers, generative models, or multimodal data; 2) Physics-guided deep learning with tighter embedding of scattering/radiative models; 3) Lightweight deployment for onboard satellite processing; and 4) Cross-domain expansion into medical imaging or environmental monitoring. Advancing "physics-informed, data-adaptive" hybrid intelligence frameworks will be pivotal for achieving robust, interpretable, and scalable hyperspectral unmixing, ultimately enhancing precision in Earth observation and beyond.