An energy-efficient obstacle-crossing control framework for quadruped robots

• We train a data-drive control strategy based on an adversarial motion prior that can control a quadrupedal robot to perform jumping maneuvers. • Based on physics, a trajectory prediction model is trained in the framework, which predicts multiple feasible and least energy-consuming obstacle-crossing trajectories by acquiring obstacle information and robot's current state. • The framework was successfully deployed on a real robot and completed the tasks of obstacle crossing and continuous jumping in the real world. The ability of a quadruped robot to cross obstacles is a crucial metric for assessing its adaptability in complex environments. Traditional control methods depend on precise physical modeling, which struggles to adapt to complex environments. Nowadays, embodied intelligence has become an important concept for describing agent as learning through environmental interactions. In recent years, techniques like deep reinforcement learning and imitation learning, designed to address interaction challenges, have achieved significant success in robot control. However, many challenges remain, including complex reward mechanism design, poor model generalization, and insufficient expression of physical laws. To this end, a novel energy-efficient obstacle-crossing control framework is developed, which combines the data-driven method of adversarial motion prior and the energy consumption knowledge of physics. This allows the quadruped robot to generate multiple feasible and lowest energy consumption trajectories according to the obstacle information and its current state, enabling it to successfully complete the obstacle crossing task. This framework introduces a novel paradigm for quadruped robot control.