Efficient co-adaptation of humanoid robot design and locomotion control using surrogate-guided optimization

Abstract

Recent advancements in reinforcement learning (RL) and computational resources have demonstrated the efficacy of data-driven methodologies for robotic locomotion control and physical design optimization, providing a scalable alternative to traditional human-crafted design paradigms. However, existing co-design approaches face a critical challenge: the computational intractability of exploring high-dimensional design spaces, exacerbated by the resource-intensive nature of policy training and candidate design evaluations. To address this limitation, we propose an efficient co-adaptation framework for humanoid robot kinematics optimization. Building on a bi-level optimization architecture that jointly optimizes mechanical designs and control policies, our method achieves computational efficiency through two synergistic strategies: (1) a universal policy generalizable across design variations, and (2) a surrogate-assisted fitness evaluation mechanism. We implement the method with humanoid robot Kuafu, and by experimental results we demonstrate the proposed method effectively reduces the cost and the optimized design can achieve near-optimal performance.