Towards a more practical data-driven biped walking control

Walking motion is a fundamental skill involved in daily bipedal activities. Generating the walking motion from mocap data requires a controllable and responsive character in both animation and robotics. The physic-based data-driven technologies nowadays have shown great value in the motion imitation of captured walking data. In this paper, we propose a new data-driven bipedal control framework by adjusting the mocap data with physics-based model to achieve a robust and natural performance while keeps the style of the data as complete as possible. In our framework, the desired bipedal action is designed from the balance strategy and a cascaded position-velocity control with gravity compensation is proposed for joint tracking control of the desired action. We demonstrate the effectiveness of this framework through different external perturbation tests. This framework more fully integrates the original data into desired-action design while keeps a robust and stable performance in the simulation.