In context
By 2025, quadruped robots had moved beyond walking to agile tasks like parkour, yet most designs relied on engineer intuition for leg dimensions and optimized control separately. This gap limited performance ceilings, prompting research into co-design—jointly optimizing mechanical structure and control—to unlock greater agility.
What was reported
Researchers from Zhejiang University, Deep Robotics, and USTC proposed a pre-training-fine-tuning framework for structure-control co-design of quadruped robots, inspired by asymmetric limbs in animals like cougars. The method optimizes thigh and shank lengths to enhance parkour capabilities such as jumping height and distance.
The framework first pre-trains a control policy across thousands of robots with varied leg lengths using spatial domain randomization in Isaac Gym, combined with discount regularization to improve generalization. In the fine-tuning stage, Bayesian optimization iteratively selects candidate structures, and the pre-trained policy is fine-tuned for each candidate to ensure near-optimal control.
Experiments showed that fine-tuning required only 400 steps—about 6.67% of conventional training—to achieve stable performance comparable to policies trained from scratch for each structure. The co-design approach significantly outperformed independent control optimization, enhancing extreme parkour performance.
Why it mattered
This work demonstrated a computationally efficient path to co-design, reducing time while maintaining control optimality. It offers a practical method for manufacturers to tailor robot morphology to specific tasks, potentially accelerating development of agile legged robots for industrial inspection, logistics, and other demanding environments.
“The co-design strategy substantially exceeds the conventional method of independently optimizing control strategies, providing an innovative approach to enhancing the extreme parkour capabilities of quadruped robots.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-09-16 · “基于预训练—微调框架的四足机器人结构—控制协同设计”