The article discusses the use of machine learning techniques for robot dynamics learning and control. The authors propose a Hamiltonian formation of a neural ordinary differential equation (ODE) network to approximate robot dynamics. This method ensures energy conservation and accounts for energy-dissipation effects such as friction and drag. The paper also discusses the development of energy shaping and damping injection control for the learned Hamiltonian dynamics. This approach enables a unified method for stabilization and trajectory tracking with various robot platforms.
Publication date: 19 Jan 2024
Project Page: https://thaipduong.github.io/LieGroupHamDL
Paper: https://arxiv.org/pdf/2401.09520
