With robotaxis, more VMT is inevitable. But perhaps, not much.

The simulations on robotaxis have been clear for years: more VMT is inevitable, because trip-ends are poorly matched to trip origins. If two people are going from home to work, that's four Origin-Destination trips, of which only two are ferrying people*. More miles on the road, but less demand for parking at destinations. 

Is that a worthwhile social trade-off? Ignoring the equity question ("Who benefits?"), and converting it all into something fungible (dollars), what's the value of a parking space versus the cost of road capacity? Answering that question requires bounding it in time and space: a parking space where, and VMT when. The value of a parking space in Manhattan could easily be in the hundreds of thousands of dollars, while a parking space in rural west Virginia might actually be negative. Likewise, the cost of adding an additional car to the roadway at 3a is basically zero - the road is sized to serve peak hour traffic, and 90% of the time, isn't carrying anything like that volume. So as a thought experiment, say it's the peak hour and parking for a major league baseball event, where parking is $300 an hour for a prime spot. For the robotaxi user, not parking is a clear win - the cost of the trip might be $45, for a $90 round trip--still a huge win. But for society, what's the marginal cost of the additional VMT generated by the robotaxi moving to its next fare.  Extreme case is that it's pure dead-head - no outbound fare from the stadium and drives all the way back to where it picked up the last guy to start its next fare. 

Quick google suggests a value of 1.6 cents per mile for federal costs, plus .1-.5 per state. Call it two cents a mile. Say the stadium trip is 20 miles. So that's .02 * 20, or forty cents. 

Ride-hailing companies will argue that this is an absurd case, but empirical evidence has not been made broadly available. Published meta-analysis suggests robotaxis will generate more VMT, but not that much. Six percent is more than a rounding error, but far less than the 50% I've seen bandied about in the past. I suspect that it's density dependent - if you convert a whole city to ride-hailing, you get much better matching--the next pickup after a drop-off, the dead-head DO pair, is much lower. As a thought experiment, consider a city with only one robotaxi, taking fares on a more or less random basis. Adding a second robotaxi substantially increases the chance that time and place can be made to match, and adding a thousand robotaxis is exponentially better, as the DO distance and time match converge exponentially. 

But as with many urban schemes, the devil is in the details. Lots of schemes falter because while the end-state is optimal, a long-term project requires every intermediary stage to be sufficiently self-supporting to sustain itself if/when there is a pause in the progress of the scheme. In a way, that speaks well for the future of robotaxis--they've proven to be financially viable even at very low levels of penetration.

Also clarifies why Uber expanded from city to city only incrementally - it required massive subsidy (to drivers and passengers) to ensure sufficient market participants for the scheme to work in the first place. Lacking drivers, wait times would be analogous to that of phone-hailed taxis. But lacking a series of passengers, no driver will find it worth their while to make that first pickup. So the whole market had to be bootstrapped into existence.