This thesis presents a way to solve the issue of a jump with a run-up over an obstacle using a quadruped robot. To be applicable in reality, the ability to run on flat terrain should not be degraded. For this purpose, a hierarchical framework with two neural networks is used in this work. Both networks are trained using reinforcement learning. The switch between the two networks is done based on the distance between the robot and the hurdle. Simulations show a high success rate for the developed policy. The possibility to transfer this approach to reality is proven by an experiment on a real robot.