I believe that self-supervised representation learning and exploration together have the potential to let deep reinforcement learning match or surpass human sample efficiency, even in challenging vision-based problems.