The marginal driver cost is what creates equilibriums in the conventional modes. Either they're scheduled to work a shift, or they're owner-operators. Both create constraints on how many vehicles can be dispatched, the form they have to take to satisfy demand, and the logistics around them(train tracks, parking spaces, automotive supply services, etc.). The large footprint of private auto and taxis lies within these constraints, as does the high-cap status of buses and trains.
The new modes are actually just the cost of the vehicle and support team, which means that the fleet can feasibly operate tiny, light, high frequency vehicles instead of or in addition to huge, consolidated, scheduled ones. You could have a self-driving three-wheeler, a self-driving minibus and so on. Loading can be standardized by policy.
That they aren't doing that right away just reflects the current cost structure and legal framework being weighted towards larger prototypes following conventional modes. But the reason why a robotaxi makes a better taxi is because it isn't a taxi in practice, it's transit. It's centrally planned, it has standardized fleet management.