Generating coordinated bounding locomotion for a small-scale quadrupedal robot with thoracic-pelvic articulation via reinforcement learning

Small-scale cursorial quadrupedal animals coordinate their spines and limbs for rapid and agile locomotion. Although small-scale quadruped robots have demonstrated impressive mobility in various unstructured environments, they rarely fully leverage the coordination between their spines and limbs, which limits the boundaries of their physical capabilities. Here, we present a hybrid adaptive control framework systematically integrating biological kinematics with reinforcement learning, utilizing pika-inspired morphology as a functional template and contextual motivation for dynamic bounding gaits via coordinated pelvic-thoracic articulation. Inspired by biological locomotion, we developed a parameterized expert model of joint angles in the pika (Ochotonidae), a small-scale cursorial animal, and employed it to generate optimized foot-end trajectories. We further explored the optimal gaits using reinforcement learning to maximize the motor performance of the generated reference trajectories on the robotic quadruped. Notably, the policy achieves coordinated motion highly correlated with biological data and exhibits stable limit-cycle dynamics. The experimental results demonstrate a substantial speed improvement over baseline reinforcement learning. This study provides valuable insights for developing more coordinated and rapid quadrupedal gaits, potentially bridging the performance gap between small-scale robotic and biological quadrupeds.