Koopman-Based Online Identification With Sim2Real Transfer for Hydrodynamic Modeling of Turtle Inspired Robot

Abstract

This letter addresses critical challenges in the dynamic modeling of biomimetic underwater robots, including heavy reliance on prior knowledge and difficulties in tracking time-varying models. To tackle these issues, we propose a control-oriented online identification framework based on the Koopman operator and successfully apply it to the hydrodynamics modeling of a sea turtle-inspired biomimetic robot. Initially, a prior model is obtained offline through simulation data based on Koopman theory. Subsequently, an online update algorithm integrated with Kalman filtering is developed to calibrate the model parameters in real time using a small amount of experimental data. Simulation and experimental results demonstrate that the proposed Koopman model outperforms baseline methods in long-term prediction accuracy and generalization capability. After online updates, the model achieves a 42.13% improvement in prediction accuracy over the baseline. In orientation tracking experiments under flow disturbances, the proposed framework enhances control accuracy by 63.4% compared to the baseline, while exhibiting smaller overshoot and stronger disturbance rejection capabilities. Overall, this study provides valuable insights for high precision dynamic modeling of biomimetic underwater robots.