Learning to walk: optimizing trajectories and policies for real robots and dynamic tasks - Naver Labs Europe