A numerical simulation method for bionic fish self-propelled swimming under control based on deep reinforcement learning

In order to simulate the under control self-propelled swimming of bionic fishes, a coupling method of hydrodynamics/kinematics/motion-control is presented in this paper. The Navier-Stokes equations in the arbitrary Lagrangian-Eulerian framework are solved in parallel based on the computational domain decomposition to simulate the unsteady flow field efficiently. The flow dynamics is coupled with the fish dynamics in an implicit way by a dual-time stepping approach. In order to discretize the computational domain during a wide range maneuver, an overset grid approach with a parallel implicit hole-cutting technique is adopted and coupled with morphing hybrid grids around the undulation body. The motion control of the fish swimming is realized by a deep reinforcement learning algorithm, which makes the fish model choose proper undulation manner according to a specific purpose. By adding random disturbances in the training process of fish swimming along a straight line, a simplified two-dimensional fish model obtains the ability to swim along a specific trajectory. Then in subsequent tests, the two-dimensional fish model is able to swim along more complex curves with obstacles. Finally, the starting process of a three-dimensional tuna-like model is simulated preliminarily to validate the ability of the coupling method for three-dimensional complex configurations. The numerical results demonstrate that this study could be used to explore the swimming mechanism of fishes in complex environments and to guide how robotic fishes can be controlled to accomplish their tasks.