Abstract

Dynamic manipulation of passively connected objects by legged-arm robots presents unique challenges due to nonlinearities, such as a floating base and the passive nature of the end effector. Although learning-based approaches offer potential for handling these complex dynamics, they often require extensive data, making them computationally demanding. On the other hand, optimization-based methods face limitations when dealing with unknown objects and adapting to variable environments. This paper proposes a hybrid control framework that selectively applies learning only to the end-effector trajectory while utilizing Whole-Body Impulse Control (WBIC) with Model Predictive Control (MPC) for overall stability and motion tracking. The proposed framework highlights the potential of RL-integrated control in reducing the learning burden and improving control efficiency and robustness in autonomous material-handling tasks in complex setting.