In 2014, the concept of combining perception and control using end-to-end reinforcement learning for robotic manipulation skills from vision was considered controversial in the robotics world but proved to be effective.