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

Area of Science:

• Robotics
• Materials Science
• Artificial Intelligence

Background:

• Magnetic soft millirobots offer untethered locomotion in narrow spaces but are difficult to design due to complex interactions.
• Current design processes are largely intuition-driven, lacking systematic optimization for real-world performance.

Purpose of the Study:

• To propose an uncertainty-aware, data-efficient inverse design methodology for magnetic soft millirobots.
• To automate the design of robots for confined-space crawling in contact-rich, non-smooth environments.

Main Methods:

• Integration of a physics-based Cosserat rod model with Gaussian Process-based Bayesian optimization.
• Incorporation of domain randomization to model contact uncertainty and mitigate sim-to-real discrepancies.
• Utilizing channel segmentation for critical geometric bottlenecks to enhance optimization efficiency and accuracy.

Main Results:

• Optimization time was halved, and R² increased by an order of magnitude in serpentine channels.
• Optimized robots demonstrated stable crawling across diverse conditions, outperforming baseline designs without failures.
• In coronary artery-mimicking geometries, optimized designs achieved 2.66 mm/s, nearly doubling baseline speeds (1.42 mm/s).

Conclusions:

• The developed methodology provides an uncertainty-aware inverse design approach for task-driven magnetic soft millirobot design.
• This work paves the way for automated design-to-deployment pipelines for real-world robotic applications.
• Optimized millirobots show enhanced reliability and performance in complex, confined environments like biological vasculature.