Q-attention: Enabling Efficient Learning for
Vision-based Robotic Manipulation