AI-Driven Magnetization Reprogramming for Soft Robotics: Learning-Based Real Time Control of In Situ Magnetic Domains

Magnetization reprogramming offers a powerful mechanism for shape control in soft magnetic robots, yet its full potential remains limited by manual tuning and lack of autonomy. This study introduces an AI-driven framework that transforms magnetization reprogramming into a predictive, adaptive, and geometry-agnostic control strategy. By learning the nonlinear mapping between magnetization profiles, structural mechanics, and magnetic-field interactions, the AI model enables real-time optimization of deformation across one-, two-, and three-dimensional architectures. Experimental results demonstrate substantial improvements in shape prediction, with the error reduced from 12.4% to 3.1% on the held-out test set, together with a reduction in AI inference time from 180 ms to 42 ms and improved robustness to fabrication and sensing variability. The system achieves contact-free navigation, 100% vascular branch entry success, sub-phase resolution in cilia arrays, and multi-face morphing in 3D structures, all under uniform magnetic fields. A conceptual framework formalizes the integration of task objectives, sensor feedback, magnetization reprogramming, and magnetic actuation into a closed-loop intelligent morphing system. Together, these findings establish AI-enhanced magnetization reprogramming as a scalable paradigm for next-generation soft robotics, enabling autonomous shape adaptation, multifunctional actuation, and high-fidelity control in constrained environments.