HNN-Transformer Integrated Network for Estimating Robot Position

This paper proposes a novel model architecture, the Hamiltonian neural network (HNN)-Transformer, which capitalizes on the strengths of both Hamiltonian neural networks and Transformers to effectively model and predict the behavior of physical systems. The proposed structure is designed to solve the problem that arises over time when predicting the state of a 2D robotic system. The HNN component of our model makes it possible to incorporate known physical laws into the learning process, while the Transformer component enables effective processing of sequentially input data. The performance of the proposed method was confirmed through simulation, and it was confirmed that this novel model is possible to leverage the strengths of both HNNs and Transformers to achieve improved performance than the existing independent HNN and Transformer on the challenging task, respectively. This suggests that the integration of physical understanding and sequence processing is a promising direction for modeling complex dynamic systems.