Abstract—Policy transfer is an efficient approach for developing specific robots. Its effectiveness depends on high-quality imitation datasets and a stable learning process. However, substantial differences in geometry and dynamics between source and target robots pose challenges. Purely kinematics-driven mapping methods and manual parameter tuning often fail to maintain kinematic-dynamic consistency. In this study, we transfer control policies from the quadruped robot Unitree Go1 to our self-developed heavy wheel-legged robot Tiangou. We propose a Consistency-Aware Retargeting (CAR) method. This extends conventional inverse kinematics by adding dynamic consistency constraints. Using motion data from Go1’s Model Predictive Controller (MPC), CAR generates a reference dataset for Tiangou. We then integrate Bayesian Optimization (BO) into the imitation learning framework. This enables autonomous tuning of policy model structures and optimization hyperparameters. Experiments show that CAR reduces foot-end position errors, mitigates joint angular velocity fluctuations, and decreases foot-end slippage. Moreover, Bayesian optimization improves sample efficiency and training stability. These contributions establish a practical foundation for policy transfer across heterogeneous robotic platforms.  

Keywords—imitation learning, motion retargeting, Bayesian optimization, policy transfer, wheel-legged robot

Cite: Chengleng Han, Lin Xu, and Changshun Huang, "Dynamics-Driven Policy Transfer to Heavy Wheel-Legged Robots via Imitation and Optimization," International Journal of Mechanical Engineering and Robotics Research, Vol. 15, No. 1, pp. 28-37, 2026. doi: 10.18178/ijmerr.15.1.28-37

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).