Google DeepMind published RT-2 vision-language-action model enabling robots to reason about and execute manipulation tasks from language instructions