Dynamics modeling is one of the foundational pillars of robotics research. It underpins the design and simulation of robotic systems, aids in the analysis of their safety and reliability, and enhances the synthesis of controllers, particularly in model-based control methods. Traditionally, modeling the dynamics of robots has relied on physics-based methods, where dynamics models are built from first principles in a bottom-up manner, providing strong generalization abilities. However, as robotic systems become more complex, it is increasingly difficult to capture their behaviors using purely physics-based approaches. Over the past decade, data-driven approaches, such as deep learning, have shown promise in tackling this complexity. Yet, these methods come with significant limitations, especially the need for large amount of training data, which is often scarce or difficult to collect for complex robotic systems. This thesis introduces Knowledge-Based Neural Ordinary Differential Equations (KNODE), a novel approach for modeling complex robot dynamics that bridges the gap between physics-based and data-driven methods. KNODE offers a hybrid modeling technique that integrates domain-specific physical knowledge into machine learning models, harnessing the strengths of both approaches. By embedding first-principles knowledge into the learning process, KNODE significantly improves model performance, reducing the data requirements and enhancing the generalization ability of learned models. This thesis will demonstrate the applications of KNODE across diverse systems, including quadrotors, soft robots, and other complex dynamical systems. Each application not only serves to validate the proposed approach but also provides insights into its practical implications and the potential for further research in robot dynamics modeling. By combining machine learning with domain-specific insights, this thesis advances the field of robotic systems modeling, providing a practical framework for handling the increasing complexity of modern robotic systems in modeling and control tasks. KNODE not only enables more accurate predictions of system behavior but also offers a path forward for reducing the reliance on large datasets, making it a valuable tool for both robotics researchers and engineers.