A Data-Driven Inverse Design Methodology for Magnetic Soft Millirobots Navigating in Confined Spaces

Magnetic soft millirobots enable untethered locomotion in narrow environments. However, their design remains largely intuition-driven due to complex interactions among soft-body deformation, magnetic actuation, and environmental contact. Here, we propose an uncertainty-aware, data-efficient inverse design methodology tailored for contact-rich, non-smooth environments. It integrates a physics-based Cosserat rod model with Gaussian Process-based Bayesian optimization to automate robot design for confined-space crawling. To mitigate sim-to-real discrepancies, domain randomization is incorporated by explicitly modeling contact uncertainty. Channel segmentation focusing on critical geometric bottlenecks enhances efficiency and accuracy, halving optimization time and increasing R² by nearly an order of magnitude in a serpentine channel. Optimized robots consistently outperform arbitrarily selected baseline designs, achieving stable crawling across heterogeneous conditions without failures such as coiling or jamming. When applied to coronary artery-mimicking geometries, the optimized design reached 2.66 mm/s, nearly doubling the average baseline speed of 1.42 mm/s. Validation in an artery-mimicking channel with 2.5 mm out-of-plane undulations further demonstrates the reliability of the optimization framework; notably, the optimized design maintained uninterrupted motion while 25% of baseline designs got stuck. This work establishes an uncertainty-aware inverse design methodology for task-driven magnetic soft millirobot design, paving the way toward automated design-to-deployment pipelines for real-world applications.