In my opinion, the approach of embodiment is the surest path to creating genuine understanding in ML systems.

The topology of the model ends up being tuned to represent the world through its interactions within the physical space.

This creates a shared context we can use to communicate with the AI. If you tell it go get a cup from the table, it will have an idea of what cup and table are similar to ours.

microsoft.com/en-us/research/g