Accelerating Soft Robot Evolution Using N-gram-based Controller Inheritance and Genetic Co-Design

Abstract

Soft robots have drawn increasing attention due to their flexible morphologies suited for complex tasks in dynamic environments. However, the co-design of structure and control for soft robots remains a challenge, primarily due to the vast search space and high computational costs. A major limitation of current methods is the inefficient reuse of knowledge across generations, causing repeated learning of similar behaviors. This paper proposes an N-gram-based controller inheritance framework, integrating genetic algorithms for structural evolution with proximal policy optimization for controller training. By capturing sequential behavioral patterns from ancestor controllers, our approach reduces redundant learning and speeds up optimization. Experimental results demonstrate that we achieve a faster convergence rate and an average 12.7% improvement in final fitness scores compared to existing methods. This study demonstrates the effectiveness of N-gram inheritance in improving soft robot co-design.