In the case of Google/DeepMind’s SIMA it is an instruction-following agent for simpler, but menial tasks in a game. It is particularly not autonomous, and has no notion of reward. And what is being used is a modified behavior cloning with a text-conditioned policy network.