Imagine robots learning soccer strategies not from lines of code, but by watching videos of real matches, just like humans do. That’s the groundbreaking idea behind LLCoach, a new system that uses the power of large language models (LLMs) to generate game plans for robot soccer teams. Traditionally, robot soccer teams operate on pre-programmed instructions. But real games are full of unpredictable moments. LLCoach tackles this challenge by using a multi-role approach. A 'coach' VLM (visual language model) analyzes video frames, identifies player positions and the ball's location, and then suggests a high-level strategy in natural language, much like a human coach explaining a play. This strategy is then refined into a detailed, executable plan for each robot on the team, using the LLM to consider the specific actions each robot can perform. These actions, such as passing or kicking, are stored in a database, allowing the LLM to construct robot-specific plans with greater accuracy, like selecting the right ‘play’ from a playbook. Remarkably, this AI coaching system significantly outperforms traditional methods. In simulated matches, robots guided by LLCoach achieved a 90% success rate in scoring goals, compared to just 30% for robots using human-written code. They also scored faster and made more passes, demonstrating a more strategic approach to the game. The implications extend far beyond robot soccer. LLCoach offers a glimpse into a future where robots learn complex tasks by observing and understanding human behavior, potentially revolutionizing how we teach and deploy robots in various fields, from manufacturing to healthcare. This research moves us closer to the 2050 RoboCup goal: robots playing soccer as skillfully as humans. The next step? Teaching robots with videos of *human* soccer matches. Just imagine the possibilities.