Imagine a world where robots learn complex skills not through painstaking programming, but by following lesson plans crafted by large language models (LLMs). This is the intriguing premise of a new research paper that introduces "CurricuLLM," a system designed to generate automatic task curricula for robots. Traditionally, teaching a robot something new, like running or manipulating objects, requires intricate code and extensive human intervention. However, CurricuLLM leverages the power of LLMs, like GPT-4, to create structured learning experiences for robots in diverse simulated environments. The process starts with the LLM devising a sequence of progressively challenging subtasks, expressed in natural language. For instance, if the ultimate goal is to make a humanoid robot run, CurricuLLM might define initial steps like "maintain stability" and "learn to walk." Subsequently, the LLM translates these language descriptions into executable code, essentially crafting reward functions and goal distributions that the robot can understand. It's like creating a set of incentives for the robot to learn and adapt. The robot then trains on each subtask, gradually building up its abilities. To make sure the learning stays on track, another LLM acts as an evaluator, assessing the robot’s performance based on its movements. This evaluator ensures the robot effectively learns from the curriculum without straying or developing undesirable behaviors. Initial experiments with CurricuLLM show promise in teaching robots manipulation, navigation, and even locomotion skills in simulation. In one challenging experiment, CurricuLLM successfully guided a simulated humanoid robot to follow complex velocity and heading commands, achieving performance comparable to a robot trained with human-designed rewards. This is exciting because manually crafting rewards takes time and expertise. CurricuLLM could automate and democratize the learning process for robots. Moreover, the researchers successfully deployed the learned policy onto a real-world humanoid robot, marking a successful leap from simulation to physical reality. While more development is needed to improve the curriculum design and integrate more nuanced performance feedback, CurricuLLM offers a compelling vision of the future. One where LLMs could play a pivotal role in shaping the skills and abilities of robots, potentially unlocking new applications across various fields.