Translation: Recently, NVIDIA introduced an AI system named Eureka, which leverages the generative capabilities of GPT-4 to train robots to perform various complex tasks, such as pen spinning and drawer opening, through reinforcement learning. Research indicates that Eureka's reward function outperforms human-designed reward functions in 83% of 29 benchmark tests, leading to over a 50% improvement in the robots' average performance, especially in handling complex tasks like high-dimensional motor control. This research offers new insights and technical approaches for the application of reinforcement learning in robotics. The project is now fully open-source.