Imagine a robot learning complex skills not through painstaking manual programming, but with the guidance of an AI coach that constantly refines its teaching methods. This is the promise of ROSKA, a groundbreaking new framework that uses large language models (LLMs) to revolutionize robot skill acquisition. Traditionally, training robots involves designing intricate reward functions that incentivize specific behaviors. This process is often complex, requiring expert knowledge and extensive trial and error. However, ROSKA simplifies this by employing LLMs to dynamically generate and refine these reward functions. Unlike existing methods that train policies from scratch with each new reward function, ROSKA evolves the reward function and the policy in tandem. This co-evolutionary approach allows the AI coach (the LLM) to adapt its 'teaching' based on the robot's current performance. The system starts with an initial set of reward functions generated by the LLM. A policy—essentially the robot's strategy—is then trained using these rewards. Instead of discarding the policy and starting over with a new reward function, ROSKA cleverly fuses the existing policy with some random exploration. This fusion allows the robot to retain previously learned skills while also staying adaptable to the refined reward function. The magic happens through a process called Short-Cut Bayesian Optimization (SC-BO). SC-BO efficiently determines the optimal blend between the robot’s learned skills and new exploration, ensuring the fastest possible learning. This entire cycle of reward generation, policy fusion, and optimization repeats, with the LLM constantly refining its reward function based on the robot's progress. The result is a significantly faster and more efficient learning process. In experiments across various challenging robotic tasks—from manipulating objects with a robotic hand to controlling a humanoid robot's locomotion—ROSKA demonstrated remarkable improvements. It used less training data than existing methods while achieving significantly better performance, sometimes even surpassing human-designed reward functions. ROSKA opens doors to a new era of robot learning, where AI coaches empower robots to master complex skills with unprecedented speed and efficiency. This technology paves the way for more adaptable robots that can learn new tasks quickly and effectively, pushing the boundaries of what's possible in robotics and automation.