This thesis addresses the challenges of model-based planning and control in legged locomotion, particularly the trade-off between computational speed and robot performance presented by different levels of model complexities. Full-order models, while rich in detail, are often too computationally demanding for real time planning, whereas conventional reduced-order models (ROMs) tend to oversimplify the dynamics, limiting overall performance potential. Our research focuses on a novel approach – the direct optimization of ROMs. This study seeks to enhance the performance of legged robots by automatically discovering the optimal ROMs that simultaneously deliver high robot performance while maintaining the necessary low dimensionality for real time planning applications. In this work, we formulate problems, provide algorithmic solutions, and deploy optimized ROMs on real robots. In the beginning of the thesis, we focus on a special case where we aim to find whole-body orientation coordinates (WBO) for legged robots that minimize angular momentum errors. This optimal WBO, while being a simple forward kinematic function, serves as a proxy of the real angular momentum and can be applied to complex tasks such as humanoid natural walking. In the second part of the thesis, we formulate a bilevel optimization problem to find optimal ROMs agnostic to controller choices, driven by user-defined objectives and task distributions. The results show substantial improvements in walking speed, ground slope adaptability and torque efficiency on a bipedal robot Cassie. Lastly, we cast the ROM optimization problem as a model- based reinforcement learning (RL) problem to further improve the model performance. This does not only show better performance improvements in experiment but also provide an easier way to implement model optimization and to realize the model performance on the robot.