<p>Industrial robotic manipulators play a central role in modern manufacturing but remain vulnerable to faults in sensors and actuators that may compromise safety and operational continuity. This paper proposes an unsupervised anomaly detection framework that integrates a high-fidelity Digital Twin (DT) of the Stäubli TX60 robot with several deep autoencoder architectures. The investigated models include a compact feedforward autoencoder (Dense AE), a deeper variant with higher representational capacity (Deep AE), and two recurrent architectures based on Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), designed to capture temporal dependencies in motion data. The DT, calibrated with CAD specifications and experimental joint trajectories, generates both nominal and faulty motion sequences, enabling controlled training and rigorous validation. All autoencoders are trained solely on fault-free data, and anomalies are detected using reconstruction error. Experiments show that the Deep AE provides the best balance between precision and recall, while the LSTM AE excels in capturing gradual temporal degradations. The Dense AE, although less accurate, offers low-latency performance suitable for real-time deployment. Comparisons with classical approaches, including Principal Component Analysis (PCA), Isolation Forest, and One-Class SVM, confirm the superiority of deep autoencoder-based methods in modeling nonlinear and multivariate joint dynamics. In addition, a Variational Autoencoder (VAE) was implemented as a representative baseline from recent deep anomaly detection methods. While the VAE slightly outperformed traditional AEs in terms of detection accuracy, it incurred higher inference latency, highlighting a trade-off between accuracy and real-time feasibility. This trade-off further motivates the choice of autoencoders in manufacturing scenarios where low-latency monitoring is critical. Finally, the framework is deployed in Simulink and validated under real-time fault injection, enabling joint-level fault localization and interpretable diagnostics. The results highlight the potential of combining Digital Twins with deep learning for scalable, interpretable, and real-time fault detection in industrial robotics, paving the way for predictive maintenance and safer robotic operations.</p>