Pretraining-Finetuning Framework for Efficient Codesign of Legged Robots

In nature, animals with exceptional locomotion abilities, such as cougars, often exhibit asymmetry between their forelimbs and hindlimbs. This observation inspires our investigation into whether optimizing leg length in legged robots can impart similar locomotion capabilities. In this article, we propose a method that cooptimizes mechanical structure and control policies to enhance the locomotion performance of legged robots. Specifically, we introduce a pretraining-finetuning framework that not only ensures optimal control strategies for each mechanical configuration but also enhances time efficiency. Furthermore, we develop an innovative training approach for our pretraining network that integrates spatial domain randomization with discount regularization techniques, significantly improving the network’s generalizability. The effectiveness of the proposed method is validated through two quadrupedal locomotion tasks and two parkour-style tasks. Experimental results demonstrate that the pretraining-finetuning framework markedly enhances overall codesign performance with reduced time expenditure. Additionally, experimental validation is performed, confirming that the cooptimized morphology and control strategy effectively mitigate the robot’s energy expenditure during designated tasks.