This paper presents a novel method for training and transferring agile and dynamic motor skills for legged robots, specifically the ANYmal robot. The approach combines classical models of rigid-body dynamics with reinforcement learning to train a neural network policy in simulation, which is then deployed on the physical robot. The method addresses the challenges of training on real robots, such as the complexity and cost of traditional control algorithms, by leveraging fast and automated data generation schemes. The trained policy enables the ANYmal robot to achieve precise and energy-efficient following of high-level body velocity commands, run faster than ever before, and recover from falls in complex configurations. The results demonstrate significant improvements over previous methods, including a 25% increase in speed and the ability to perform dynamic recovery from falls. The approach is computationally efficient, robust to hardware changes, and applicable to a wide range of rigid body systems.This paper presents a novel method for training and transferring agile and dynamic motor skills for legged robots, specifically the ANYmal robot. The approach combines classical models of rigid-body dynamics with reinforcement learning to train a neural network policy in simulation, which is then deployed on the physical robot. The method addresses the challenges of training on real robots, such as the complexity and cost of traditional control algorithms, by leveraging fast and automated data generation schemes. The trained policy enables the ANYmal robot to achieve precise and energy-efficient following of high-level body velocity commands, run faster than ever before, and recover from falls in complex configurations. The results demonstrate significant improvements over previous methods, including a 25% increase in speed and the ability to perform dynamic recovery from falls. The approach is computationally efficient, robust to hardware changes, and applicable to a wide range of rigid body systems.