Neuroevolution of Decentralized Decision-Making in N-Bead Swimmers Leads to Scalable and Robust Collective Locomotion

Many microorganisms are capable of swimming through viscous fluids such as water in order to search for nutrients, swim toward oxygen or light, or to escape from predators. To navigate their environment they often perform large nonreciprocal periodic deformations of their shape, by waving appendages such as cilia or flagella, or by deforming their entire body. Even unicellular organisms are fundamentally made of parts, which need to be cooperatively utilized to allow these creatures to navigate their environment, without using a centralized control mechanism. Here, we investigate the physical implications of decentralized decision-making of the actuators of a generalized N-bead Najafi Golestanian microswimmer, self-propelling via coordinated non-reciprocal swimming strokes. We treat each bead as an artificial neural network-based agent that perceives information about its neighbors and whose actions induce strokes of its adjacent arms. With neuroevolution techniques, we evolve optimal policies for the single-bead decision centers such that the N-bead collective efficiently self-propels as an individual, allowing us to investigate optimal locomotion policies for increasingly large microswimmer bodies. We demonstrate that such decentralized policies are robust and tolerant concerning morphological changes or defects and facilitate cargo transport or drug delivery applications "out of the box", without further optimization. Our approach allows us to train large swimmers ($N=100$ and more), and we show that long-wavelength solutions lead to surprisingly efficient swimming gaits. Our work is of relevance to understand robust locomotion of biological microswimmers, to develop robust artificial microswimmer navigation strategies, and, in a broader conceptional context, for Artificial Life< and in general emergent levels of individuality.