Wall detection for a microswimmer in Stokesian flows by using hydrodynamic signals

Perceiving the surrounding environment via hydrodynamic signals is a critical survival skill for numerous aquatic organisms and represents a promising technology for intelligent robotic platforms. This study investigates the wall-detection potential of a moving microswimmer leveraging self-induced flow signals. The problem is idealized as a spherical microswimmer undergoing constant-velocity translation or settling in the vicinity of an infinite wall within the Stokes regime, with surface-mounted sensors enabling the perception of flow signals. Analytical solutions are derived to construct comprehensive datasets for neural network training. Results demonstrate that a fixed-attitude microswimmer can accurately predict its height from the wall and translational direction using pressure and shear stress magnitude collected by only a single nadir sensor. Notably, integrating both signals is preferable for avoiding blind zones, and increasing redundant signal inputs enhances robustness against noise. For variable-attitude scenarios, the multilayer perceptron (MLP) exhibits poor generalization, whereas the graph attention network (GAT) demonstrates superior performance. Specifically, the microswimmer equipped with neural networks achieves a relative RMSE of height prediction below 5% for constant-velocity translation at heights up to 20 body radii and below 10% for settling within 8 body radii, even in noisy environments. The attention mechanism inherent in GAT facilitates the recognition of topological similarities of the input graph and exploits spatial sensor proximity to resolve complex sensing tasks. This paper provides insights into biological mechanosensing and proposes a promising approach for the autonomous navigation of microrobots via flow signals.