This paper presents a method for learning and transferring agile and dynamic motor skills for complex legged systems, such as the ANYmal robot. The approach combines simulation and real-world deployment to train a neural network policy that enables the robot to perform advanced locomotion tasks. The method involves creating a hybrid simulator that accurately models the robot's dynamics and actuator behavior, allowing for efficient training and transfer to the physical system. The policy is trained in simulation using reinforcement learning, and then deployed on the physical robot to achieve high-speed locomotion, precise control, and dynamic recovery from falls. The ANYmal robot, a medium-sized quadrupedal system, is used as a testbed for this approach. The robot is equipped with 12 Series-Elastic Actuators (SEAs), which are challenging to model due to their complex dynamics. The method addresses this by learning the actuator dynamics through self-supervised learning, enabling accurate modeling of the robot's behavior in simulation. The hybrid simulator, which includes a rigid-body simulation and the learned actuator dynamics, runs at a high speed, allowing for efficient training and testing. The trained policy enables the ANYmal robot to follow high-level body velocity commands with high precision and energy efficiency, run faster than previous records, and recover from falls in complex configurations. The policy is deployed directly on the physical robot, demonstrating the effectiveness of the approach in real-world scenarios. The method is computationally efficient, requiring minimal resources for inference, and allows for rapid adaptation to different tasks by simply changing the policy parameters. The results show that the learned controller outperforms existing methods in terms of speed, energy efficiency, and recovery capabilities. The controller is robust to changes in hardware and can handle a wide range of tasks, including high-speed locomotion and dynamic recovery from falls. The approach is applicable to a variety of legged systems and offers a promising solution for autonomous control of complex robotic systems.This paper presents a method for learning and transferring agile and dynamic motor skills for complex legged systems, such as the ANYmal robot. The approach combines simulation and real-world deployment to train a neural network policy that enables the robot to perform advanced locomotion tasks. The method involves creating a hybrid simulator that accurately models the robot's dynamics and actuator behavior, allowing for efficient training and transfer to the physical system. The policy is trained in simulation using reinforcement learning, and then deployed on the physical robot to achieve high-speed locomotion, precise control, and dynamic recovery from falls. The ANYmal robot, a medium-sized quadrupedal system, is used as a testbed for this approach. The robot is equipped with 12 Series-Elastic Actuators (SEAs), which are challenging to model due to their complex dynamics. The method addresses this by learning the actuator dynamics through self-supervised learning, enabling accurate modeling of the robot's behavior in simulation. The hybrid simulator, which includes a rigid-body simulation and the learned actuator dynamics, runs at a high speed, allowing for efficient training and testing. The trained policy enables the ANYmal robot to follow high-level body velocity commands with high precision and energy efficiency, run faster than previous records, and recover from falls in complex configurations. The policy is deployed directly on the physical robot, demonstrating the effectiveness of the approach in real-world scenarios. The method is computationally efficient, requiring minimal resources for inference, and allows for rapid adaptation to different tasks by simply changing the policy parameters. The results show that the learned controller outperforms existing methods in terms of speed, energy efficiency, and recovery capabilities. The controller is robust to changes in hardware and can handle a wide range of tasks, including high-speed locomotion and dynamic recovery from falls. The approach is applicable to a variety of legged systems and offers a promising solution for autonomous control of complex robotic systems.