Safe reinforcement learning of dynamic high-dimensional robotic tasks: navigation, manipulation, interaction