Escape from the predator-induced flow: smart prey strategies with steering and swimming actions

Plankton can sense fluid signals generated by predators and escape from them. This study explores the escape strategies using reinforcement learning (RL). The predator, modeled as a squirmer, generates flows to capture the prey entering its ciliary band. The squirmer mode characterizes the generated flow, and the higher mode corresponds to more and smaller-scale vortices. The motions of prey swimmers in the squirmer-induced flow are obtained using a Lagrangian point-particle approach. To understand the effects of different prey actions, we examine strategies obtained via Q-learning in three cases, i.e. the swimmers can only steer, only change swimming speeds, or take both actions, respectively. We compose the forth strategy where swimmers determine steering and swimming speeds separately according to the steering-only and swimming-only strategies. We find that the four strategies are all effective in the escape task. The steering-only strategy outperforms the swimming-only strategy, and strategies with both actions surpass those with only one action. The composed strategy surpasses the steering & swimming one in the training flow field. Furthermore, the robustness of strategies to swimmer shapes and flow modes is examined. All strategies are robust against various swimmer elongations. The steering-only strategy is robust in high-mode flows, whereas the swimming-only strategy is robust in low-mode flows. The steering & swimming strategy is robust for all modes, while the composed one is not robust in high-mode flows. This study investigates the possible strategies for plankton to escape from predators, revealing the effectiveness of RL in agent navigation in fluid flows.