[2E6-GS-8] Robot and real worlds:

[2E6-GS-8-01] Modeling and Control of Port-Hamiltonian Dynamics by Deep Learning

〇Tomoya Yoshioka¹, Haohui Jia¹, Takashi Matsubara¹ (1. Osaka University)

Keywords:Deep Learning, Dynamical System, Optimal Control

Identifying the dynamics of a physical system from observations is essential for precise control. In recent years, neural networks have got attention as powerful tools for representing mathematical models, and many researchers have explored methods that incorporate appropriate inductive biases into deep learning models for cost-effective dynamics identification.However, in such studies, the goal is often the identification or prediction of dynamics, and less studies are aimed at control.In this paper, we added a control mechanism to the model identifying the Port-Hamiltonian system and performed optimal control of the system using a method called AI Pontryagin.The results show that the model with physical inductive biases enables more efficient learning and control of dynamics compared to standard neural networks.