Design of Small Scale Magnetic Soft Robots Using CMA-ES and Material Point Method

Abstract

To address this challenge, this study presents an AI-driven optimization that utilizes Covariance Matrix Adaptation Evolution Strategy (CMA-ES) combined with a Material Point Method (MPM) based simulation environment, to successfully design and optimize the magnetic profile of a strip-shaped walking robot, and outperform the human-designed magnetization profile as shown by the increase in the robot's horizontal speed by 45.5%.

Additionally, we compared the CMA-ES with the Genetic Algorithm (GA), which is a widely used evolutionary optimization method. The result demonstrates that CMA-ES significantly outperforms GA in terms of convergence speed. While CMA-ES designed robots significantly outperformed the human-designed configuration, GA was unable to come close to the original design performance. This performance gap is likely because CMA-ES dynamically adapts its search distribution by using a covariance matrix. This allows for more efficient exploration of design space and increases the performance of this algorithm in solving complex problems, while GA relies on fixed mutation and crossover strategies that limit its ability.

These findings highlight the potential of AI-driven optimization in the design of the robots, enabling more innovative and efficient designs. Future work can explore multi-objective optimization to balance competing goals such as maximizing locomotion speed, improving motion accuracy, and minimizing manufacturing complexity. Additionally, this approach can be applied to different robots and locomotion modes. Extra experimental validation involving physical fabrication and testing can be done to confirm the effectiveness of the optimized designs in real-world conditions. This research represents a small step towards a fully automated design process for a magnetic soft robot.