Gathering the necessary training data for embodied models is going to be a real doozy. Hard to scale, unless you figure out how to make a perfect simulator or possibly collect data in a decentralized manner...?
1) tweak something like that to increase the likelihood of certain situations (abundance of a particular resource/object; frequent geographical feature), and
2) instruct your embodied AI to control a "player" model (to whatever degree of accuracy in articulation/mobility) to wander and perform certain types of tasks.
