The application of machine learning for modeling complex dynamic systems, such as excavators, is gaining momentum as it offers flexibility beyond traditional mathematical models. Recent advances leverage neural networks to create data-driven models that can handle the non-linear and intricate dynamics of real machinery. However, these models often depend on expensive sensors and controlled environments. This study presents a cost-effective approach to modeling the dynamics of a weakly-instrumented 25-ton CAT 323 excavator arm using stacked-input Long Short-Term Memory (LSTM) networks. We evaluate the performance of Multi-Layer Perceptron (MLP) and LSTM architectures, both with and without input stacking, to accurately simulate excavator arm motion. Our results show that combining LSTM with stacked inputs significantly improves the model’s predictive capabilities, challenging the notion that LSTM and input stacking are redundant. These findings highlight the potential of data-driven neural network models to provide accurate and efficient solutions for dynamics modeling in complex, real-world settings, paving the way for advanced AI-based strategies in the earthworks and construction industries.