Generating precise movements for advanced robots, whether for walking or manipulation tasks, poses significant challenges. While recent advancements in legged locomotion suggest we may be nearing a breakthrough, the ultimate solution remains undefined. Currently, two main strategies are being pursued to tackle this challenge: predictive control, which involves optimizing predictions of the robot movements in the immediate future while in motion, and reinforcement learning, which optimizes control policies offline. In our discussion, we will report recent results along both approaches and propose that the future lies in merging these paradigms into a unified computational method. This integrated approach would leverage offline learning to inform online adaptation, offering both efficiency and adaptability. Additionally, we will highlight the crucial role of advancements in robot hardware and emphasize the potential for artificial intelligence to further optimize the design of these robots.