Companionship with lower maintenance cost. The diversity of behavior is the problem but perhaps with the advances in neural nets we will get the behavior people desire in pets:

* unpredictable behavior within boundaries

* possible to alter behavior through repeated action (training)

* diversity of response to novel stimuli

Ultimately, what is it about a dog that makes it lovable? If we can replicate that and exploit the human desire to personify and love anything, we could create pets with intermediate intelligence and I think many people would enjoy that.

The actual product risk, I think, is making the robotic pet too useful. I suspect that would transform the pet into a tool. For instance, a robotic pet that reacts strangely and hides before the human owner detects a thunderstorm is far preferable to one which (using its link to the Internet) says “Thunderstorm approaching; batten down storm shutters”

The imperfection of understanding and the fact that the pet is a net cost center are two things that I think are crucial to mass market robotic pet adoption among those who desire companionship.