In recent years, reinforcement learning has significantly accelerated the development of legged robot control systems. The prevalent paradigm involves conducting reinforcement learning training in simulated environments initially, followed by a transition to real-world applications, a process known as sim-to-real transfer. However, this paradigm still cannot fully bridge the gap between simulation and reality. To further narrow the gap between simulation and reality, this paper proposes an innovative online learning strategy that aims to conduct training directly on the physical robot. To achieve this, we harness the power of pre-training and instruction learning to enhance learning efficiency. Additionally, we have designed an autonomous resetting system that enables the robot to automatically reconfigure and seamlessly resume learning after a fall, ensuring continuous progress. Our findings indicate that the performance of the robot after online learning has been enhanced to a certain extent compared to direct deployment using sim-to-real. The research results demonstrate the effectiveness of the Learn-in-Real paradigm in enhancing the locomotion capabilities of legged robots and provide a promising pathway for improving the performance of other legged robots.