Imagine a robot learning to navigate the world not through explicit programming, but by generating its own practice exercises. That's the intriguing idea behind Agentic Skill Discovery (ASD), a new research project that empowers robots to learn by proposing their own tasks. Traditionally, robots learn through painstaking human programming or extensive trial-and-error. ASD flips the script by using a large language model (LLM) to suggest tasks relevant to the robot's environment. For instance, in a table-top setting with blocks and a drawer, the LLM might propose tasks like "reach for the blue block," "open the drawer," or "stack the red block on the green block." Once a task is proposed, the system uses reinforcement learning (RL) to figure out *how* to achieve it. The LLM even designs the reward functions that guide the RL process, essentially defining what constitutes success for each task. To ensure the robot isn't learning nonsensical behaviors, a separate vision-language model acts as a referee. It observes the robot's actions and judges whether the task was truly completed. This feedback loop helps refine the learning process and prevents the robot from developing bad habits. The results are promising. Starting with zero skills, the robot gradually builds a library of increasingly complex actions. It learns to reach, grasp, open drawers, and even stack objects. ASD also tackles more complex tasks by breaking them down into smaller, manageable steps. For example, the task "put the blue block in the drawer" might be decomposed into "open the drawer," "pick up the blue block," and "place the block inside." This approach allows the robot to learn complex sequences of actions by building upon its existing skills. While still in its early stages, ASD offers a glimpse into a future where robots can learn autonomously, adapting to new environments and tasks with minimal human intervention. This could revolutionize robotics, enabling robots to perform a wider range of tasks in more complex environments, all while reducing the need for laborious programming. However, challenges remain. The system sometimes proposes nonsensical tasks or misinterprets its environment, highlighting the need for further refinement. Despite these limitations, ASD represents a significant step towards more autonomous and adaptable robots, paving the way for a future where AI can truly teach itself to be a robot.