Data-driven dynamic modeling for precise trajectory tracking of a bio-inspired robotic fish

We propose utilizing an attention mechanism and deep neural networks to develop a hydrodynamic identification model, integrated with a time-triggered nonlinear model predictive controller (ENMPC) for precise trajectory tracking of a robotic fish. A central pattern generator (CPG) network was employed to design a synergistic gait controller for the robotic fish that could coordinate its pectoral fins and flexible body/caudal fins to enable multimodal motion. We derived a nonlinear map between the driving parameters and the thrust/torque of the robotic fish using a computational fluid dynamics (CFD) simulation dataset. The attention mechanism was applied to incorporate laminar flow effects and construct a hydrodynamic identification model based on a bidirectional long short-term memory (Bi-LSTM) network. This identification model serves as the foundation for learning a control transformation model that operates as its inverse. Finally, event-triggered nonlinear model predictive constraints were adjusted to account for external disturbances and thereby ensure the convergence of robotic fish tracking errors while minimizing computational costs.