This is a pretty interesting paper idea, where they only use the sequence of object states as input, and generate human motion from it.

It’s used as a dataset for the TWIST paper.