DeepMind and Stanford’s new robot control model follow instructions from sketches
Sketches have rich spatial information to help the robot carry out its tasks without getting confused by the clutter of realistic images or the ambiguity of natural language instructions.