Text2Reward is a data-free framework using large language models to automate the creation of dense reward functions in reinforcement learning from natural language goals. It outperforms expert-written codes in multiple benchmarks and supports real-world deployment and human feedback refinement.