DeepMind's robotics transformer version 2 is a large language model that is trained on not just images and text, but also coordinate data of a robot's movement in space.