Learning of Sub-optimal Gait Controllers for Magnetic Walking Soft Millirobots

Area of Science:

• Robotics
• Soft Robotics
• Bioengineering

Background:

• Untethered small-scale soft robots offer potential for minimally invasive surgery and targeted drug delivery.
• Controlling these robots is challenging due to nonlinear kinematics, fabrication variability, and environmental changes.
• Adaptive control is essential for reliable performance in dynamic medical environments.

Purpose of the Study:

• To develop a data-efficient probabilistic learning approach for controlling millimeter-scale magnetic walking soft robots.
• To enable adaptive control that accounts for fabrication variability and changing surface conditions.
• To optimize the stride length performance of soft millirobots.

Main Methods:

• Utilized Bayesian optimization (BO) and Gaussian processes (GPs) for a probabilistic learning framework.
• Implemented a data-efficient learning scheme requiring minimal physical experiments.
• Demonstrated adaptation to variations in robot fabrication and walking surface roughness.

Main Results:

• Successfully adapted controller parameters for three different robots with varying fabrication characteristics.
• Showcased robot performance adaptation to walking surfaces of differing roughness.
• Achieved improved learning efficiency through knowledge transfer between robots.

Conclusions:

• The proposed probabilistic learning approach enables effective and adaptive control of small-scale soft robots.
• This method addresses key challenges in controlling soft millirobots for delicate applications.
• The demonstrated adaptability and learning transfer offer a pathway for robust robotic systems in bioengineering and medicine.