Learning Soft Millirobot Multimodal Locomotion with Sim-to-Real Transfer.

Area of Science:

• Robotics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Magnetic soft millirobots offer wireless, multimodal locomotion for minimally invasive medical applications.
• Current locomotion adaptation relies on manual signal tuning, limiting efficiency and adaptability.

Purpose of the Study:

• To develop a learning-based framework for adaptive multimodal locomotion in magnetic soft millirobots.
• To enable autonomous adaptation to diverse and unknown environments using sim-to-real transfer.

Main Methods:

• A data-driven simulation environment was created for learning magnetic actuation signals.
• Bayesian optimization and Gaussian processes facilitated sim-to-real transfer of locomotion strategies.
• Kullback-Leibler divergence was used for automated domain recognition and adaptation.

Main Results:

• The framework successfully learned periodic magnetic actuation signals for millirobot locomotion in simulation.
• Learned strategies were effectively deployed to real-world robots.
• Automated adaptation to unknown environments was demonstrated, showcasing rapid and continuous environmental response.

Conclusions:

• The proposed learning-based framework enhances magnetic soft millirobot adaptability and efficiency.
• Sim-to-real transfer and probabilistic methods enable robust locomotion in changing environments.
• This approach minimizes experimental costs and explores novel actuation solutions.