Lamarckian Co-design of Soft Robots via Transfer Learning

Abstract

In the realm of robot design, co-design aims to optimize both the structure and the controller of a robot concurrently. One approach integrates genetic algorithms to optimize the soft robot's structure with deep reinforcement learning for the controller. A significant challenge in this approach is the inheritance of the controller due to the mismatch of the sensors and actuators of the robots across generations. In this study, we propose a Lamarckian co-design method to inherit the controller optimized by deep reinforcement learning through transfer learning. In experimental evaluations through the Evogym benchmark, we demonstrate that our proposed method achieves an average reduction of 41.7% in the optimization time for robots compared to existing methods and concurrently leads to an average performance improvement of 118.5%. Furthermore, we show that combining the inheritance of controllers with the crossover of structure genomes from two robots allows for additional reductions in optimization time and improvements in performance in several tasks.