RL-GPT is a novel framework that integrates Large Language Models (LLMs) and Reinforcement Learning (RL) to enhance LLMs' capabilities in complex, embodied environments. The framework employs a two-level hierarchical structure, consisting of a slow agent and a fast agent. The slow agent is responsible for decomposing tasks into sub-actions and determining which actions can be directly coded, while the fast agent generates code and instantiates RL configurations for low-level execution. This decomposition allows each agent to focus on specific tasks, improving overall efficiency. RL-GPT outperforms traditional RL methods and existing GPT agents, achieving state-of-the-art performance in tasks such as the ObtainDiamond challenge in Minecraft, where it rapidly obtains diamonds within a single day using an RTX3090 GPU. The framework also demonstrates superior efficiency in other MineDojo tasks. RL-GPT introduces a two-loop iteration mechanism to optimize both the slow and fast agents, enabling continuous refinement of their performance. The integration of LLMs and RL allows for more efficient task learning, as LLMs can generate high-level actions while RL handles low-level execution. The framework's ability to combine coding and learning processes makes it highly effective in complex, open-world environments.RL-GPT is a novel framework that integrates Large Language Models (LLMs) and Reinforcement Learning (RL) to enhance LLMs' capabilities in complex, embodied environments. The framework employs a two-level hierarchical structure, consisting of a slow agent and a fast agent. The slow agent is responsible for decomposing tasks into sub-actions and determining which actions can be directly coded, while the fast agent generates code and instantiates RL configurations for low-level execution. This decomposition allows each agent to focus on specific tasks, improving overall efficiency. RL-GPT outperforms traditional RL methods and existing GPT agents, achieving state-of-the-art performance in tasks such as the ObtainDiamond challenge in Minecraft, where it rapidly obtains diamonds within a single day using an RTX3090 GPU. The framework also demonstrates superior efficiency in other MineDojo tasks. RL-GPT introduces a two-loop iteration mechanism to optimize both the slow and fast agents, enabling continuous refinement of their performance. The integration of LLMs and RL allows for more efficient task learning, as LLMs can generate high-level actions while RL handles low-level execution. The framework's ability to combine coding and learning processes makes it highly effective in complex, open-world environments.