Abstract: Legged robots show great potential for high-dynamic motions in continuous interaction with the physical environment, yet achieving animal-like agility remains signifcant challenges. Legged animals usually predict and plan their next locomotion by combining high-dimensional information from proprioception and exteroception, and adjust the stifness of the body’s skeletal muscle system to adapt to the current environment. Traditional control methods have limitations in handling highdimensional state information or complex robot motion that are difcult to plan manually, and Deep Reinforcement Learning (DRL) algorithms provide new solutions to robot motioncontrol problems. Inspired by biomimetics theory, we propose a perception-driven high-dynamic jump adaptive learning algorithm by combining DRL algorithms with Virtual Model Control (VMC) method. The robot will be fully trained in simulation to explore its motion potential by learning the factors related to continuous jumping while knowing its real-time jumping height. The policy trained in simulation is successfully deployed on the bio-inspired single-legged robot testing platform without further adjustments. Experimental results show that the robot can achieve continuous and ideal vertical jumping motion through simple training

Key words: Deep reinforcement learning , · High-dynamic jump , · Perception driven , · Single-legged robot