Abstract: Stability is of great significance in the theoretical framework of biped locomotion. Real-time control and walking patterns planning are on the premise that the robot works in the stable condition. In this paper, we address the crucial issue of the locomotion stability based on the modified Poincare return map and the hybrid automata. Not akin to the traditional stability criteria, i.e., the Zero Moment Point (ZMP) and the Center of Mass (CoM), the modified Poincare return map is more appropriate for both dynamic walking and non-periodic walking. Moreover, a novel high-level reinforcement learning methodology, so-called active PI2 CMA-ES, is proposed in this paper to plan the exoskeleton locomotion. The proposed learning methodology demonstrates that the locomotion of the exoskeleton is asymp-totically stable according to the modified Poincare return map criterion. Finally, the proposed learning methodology is tested by the Lower Extremity Augmentation Device (LEAD) and its effectiveness is verified by the experiments.

Key words: biped locomotion, dynamic walking, Dynamic Movement Primitives (DMP), stability, exoskeleton, bioinspiration