Large language models (LLMs) excel at various tasks, but they often struggle with physical reasoning and robotics. Imagine an LLM trying to navigate a robot through a crowded room – the abstract knowledge it possesses doesn't translate easily into real-world actions. This is because LLMs lack the 'grounding' or direct experience with the physical world. A new research project called GLIMO (Grounding Large language model with Imperfect world MOdel) aims to bridge this gap by using 'imperfect' world models, like simulators, as training grounds for LLMs. Instead of relying on perfect simulations, which can be costly and difficult to create, GLIMO uses simpler, proxy environments. Think of it like a robot learning to walk in a video game before tackling the real world. GLIMO has a clever trick up its sleeve: an LLM agent that acts as a virtual teacher. This agent explores the simulated environment, generating training data in a question-and-answer format. It iteratively refines its understanding, reflects on past experiences, and even considers hypothetical scenarios, all while creating a rich dataset for the main LLM to learn from. This approach allows the LLM to grasp the nuances of the physical world, such as the consequences of actions and environmental constraints. The results are promising. When tested on a 2D puzzle game and an urban driving simulator, GLIMO significantly boosted the performance of open-source LLMs like LLaMA. In fact, the enhanced LLMs even outperformed larger, closed-source models like GPT-4 in some tasks. This research opens exciting possibilities for robotics and AI. By grounding LLMs in simulated environments, we can equip them with the physical reasoning skills needed for real-world tasks. While the current research focuses on simulated worlds, future work aims to extend GLIMO to multimodal LLMs that can process visual and other sensory information, bringing us closer to truly intelligent robots that can understand and interact with the world around them.