Let Them Linger? The Surprising Case for Robotaxis at the Curb
An interview with the researchers behind Staging at the Curb
Where does a driverless robotaxi go between rides?
A new study — Staging at the Curb: Evaluating the impacts of shared automated vehicle fleet operations under curb usage restrictions — finds that letting AVs consume curb space might actually be the optimal outcome.
Researchers modeled 1,700 shared automated vehicles serving 68,000 daily San Francisco trips. Tying together Autofleet’s commercial fleet simulator, synthetic ridehail demand from Replica and curb availability data from INRIX, they compared several operating models: continuous circling, staging only at free non-residential curbs, access to all legally available curbs and overnight staging.
A city’s restrictions can have a big impact on VMT and deadheading. In the study’s circling-only scenario, the fleet traveled 564,650 kilometers (350,857 miles) per day. Let those robotaxis stage at the curb, and VMT falls 60%, with deadheading way down as well. Other policy levers, like restricting robotaxis to free spaces in non-residential neighborhoods, or banning overnight staging, also nudge VMT upward.
Earlier research has shown that without the incentives and emotions tied to human drivers, AVs may be apt to drive slowly between paid rides, get bogged down as passengers waste time loading into the vehicle, or have otherwise unanticipated impacts on congestion. (Reilly Brennan calls this the “pauking problem.”) I chatted with the study’s illustrious group of researchers to better understand what their research found, and see how cities can best manage their streets and curbs to make the most of the coming wave of autonomous vehicles.
Jonah Bliss: Set the stage for us; what exactly is “staging” and what does it look like on an actual city street?
Michael Hyland, Susan Shaheen, Younghun Bahk, Brooke Wolfe & Adam Cohen: We say that vehicles are “staging” when they park temporarily after dropping off one rider and are waiting to be assigned to a new ride request. More colloquially, staging essentially means resting at the curb between trips.
A staging AV would look similar to a taxi or Uber vehicle parked at the curb, except without a human driver. The AV would still be “on” and consuming some energy, but not as much energy as if it were cruising on local streets.
Give us the short version of your experiment. What kind of San Francisco robotaxi fleet did you simulate, and what different rules did you test?
We ran dozens of scenarios, but the main comparison was between fleets of AVs that (i) were not allowed to use the curb to stage, (ii) always used the curb closest to the most recent rider drop-off location to stage, and (iii) strategically relocated and staged at curbs to best serve expected future ride requests.
We simulated the second-by-second dynamics of a fleet of 1,700 robotaxis providing service on streets within the city of San Francisco. The fleet provides ridehailing service, in which a passenger would not share an AV with a stranger.
In addition to the main scenarios mentioned above, we also compared alternatives by varying whether AVs can use (i) paid or residential curb fronts, and (ii) the curb to stage at night.
Your headline finding is that preventing robotaxis from staging could increase their daily travel by roughly 60%. Where do all those extra miles come from?
Over the course of a day, there are small-to-medium-sized variations in demand that occur over relatively short periods of time (i.e., 5-30 minutes) in any passenger transportation system. As such, in a particular area of a city, a dip in the number of ride requests, especially if the dip is unexpected, will lead to in-service AVs without rides to serve. When AVs are unable to stage at the curb, they must continue circulating on nearby streets while waiting for ride requests. This behavior, commonly referred to as cruising, can substantially increase vehicle miles traveled (VMT) for AV fleets providing ridehailing service.
Did those results surprise you? Earlier research, like Adam Millard-Ball “The autonomous vehicle parking problem” have certainly suggested that robotaxi behavior, especially when not occupied by paid passengers, could have cascading congestion effects.
We expected the “no staging at the curb” scenario to increase VMT, but the large magnitude of the increase in VMT was surprising.
Adam’s study focused on AVs that are owned by individual households/persons rather than ones that are in a fleet-based service. However, one of the central insights from this study also applies to fleet-based AVs, as well as human-driven taxis and ridehailing vehicles. Since AVs do not need to park between trips, they may instead cruise while waiting for their next passenger. If the cost of parking or staging exceeds the cost of cruising, operators are likely to choose cruising.
Why do you anticipate that vehicles would not simply return to a depot, garage or private lot between rides? How might those alternative change the results?
On one hand, we would have liked to run scenarios where the only staging location for AVs was an off-street parking lot or depot between rides to stage, but we did not have the resources to run more simulations.
On the other hand, we do not think this strategy would have worked very well in terms of mitigating VMT. As we understand it, robotaxi companies do not own or have access to many off-street lots in the cities where they operate, particularly during the day. As a result, AVs would need to travel long distances from one passenger’s drop-off location to a depot and then from the depot to the next passenger’s pickup location.
To be clear, and this is one of the methodological contributions of our study, we implement “strategic defleeting,” also called “supply shaping,” in the simulation model. Supply shaping refers to a fleet management strategy where the manager determines the number of in-service or active vehicles at any time of day, based on the expected demand. As such, during periods of low demand, the fleet manager may have only 60% or even 20% (e.g., 3 am) of the AV fleet in service, while the remaining AVs are charging, being cleaned, or are turned off and not consuming energy in an off-street lot. By implementing supply shaping, the simulation does not have an unnecessary number of AVs cruising around accumulating VMT when demand is below the peak periods.
You also tested some moderate restrictions, such as keeping vehicles away from residential and metered curbs. What did those more measured policies tell you?
Preventing AVs from staging at metered curbs and curbs in residential areas will increase VMT (5.4% more empty VMT) and, by definition, prevent AVs from occupying curb space in these areas.
The effect on passenger wait times was fairly small. Does that mean operators, riders and cities all have an interest in letting vehicles wait rather than circle?
Passengers, at the margin, should be indifferent or slightly prefer cruising/circling, as our model suggests cruising AVs slightly reduce wait times. We say “at the margin” because if AV fleets scale up, all riders will be worse off if the AVs are all cruising between rides and worsening congestion.
Our results indicate that AV fleet operators and cities should be aligned in terms of avoiding the scenario where AVs are cruising rather than staging at the curb. Both entities likely want to reduce VMT, and permitting AVs to stage at the curb is an effective means to reduce VMT.
You describe robotaxis as potentially much more “productive” curb users than privately parked cars. How do you define productivity, and why is that the right comparison?
We define curb productivity as the number of people being dropped off at the curb per unit of time robotaxis spend (staging or picking up and dropping off travelers) at the curb.
Curb productivity may not be a perfect metric for comparing curb use, but it provides a useful basis for evaluation. The metric’s rationale is as follows: Cities are centers of economic, social, and cultural activity. The transportation system allows people to participate in urban activities. The curb serves as the interface between the transportation and the activity systems. Consequently, the number of people and goods crossing this interface per unit of time (i.e., curb productivity) provides a good indicator of the extent to which the transportation system facilitates urban activity.
The curb productivity metric indicates that city buses use curb space far more productively than personal vehicles parked at the curb. Similarly, based on our modeling results and observed curbside parking data from San Francisco, robotaxis use curb space approximately eight times more productively than parked personal vehicles.
Your model assumes operators want to avoid empty mileage. But if curb access becomes expensive while driving empty remains free, could cruising once again become the cheaper option?
Yes, that is certainly possible. Consequently, cities need to clearly define their policy objectives and identify the instruments best suited to achieve them. As we and others have noted elsewhere, the effectiveness of parking pricing as a travel demand management tool is likely to decline substantially as AVs become more prevalent. If cities aim to reduce VMT, congestion, and emissions in dense urban areas, they will likely need to price both curb occupancy and road use during periods of high demand. Such pricing policies should apply to all vehicle types, including AVs and human-driven vehicles, as well as both fleet-operated and privately owned vehicles.
Does this suggest that cities should price curb staging and zero-passenger travel together, rather than regulating parking and driving as separate activities?
Cities seeking to reduce congestion, VMT, and emissions should consider pricing curb occupancy and roadway use as complementary policy tools. Pricing only AVs–or only zero-occupancy AVs–for road use is unlikely to achieve these objectives, as all vehicles contribute to congestion. Policies that only target a single auto-based mode may encourage travelers to shift to other auto-based modes rather than reducing overall vehicle travel. A comprehensive pricing strategy that applies to all vehicles, including AVs and human-driven vehicles, as well as both fleet-operated and privately owned vehicles, is therefore more likely to achieve these policy goals. Such a strategy could also improve bus operating speeds and reduce public transit operating costs by alleviating roadway congestion.
Younghun Bahk adds: I mostly agree, but further down the road, cities could consider pricing zero-occupancy AV travel. Because zero-occupancy AVs have a different travel cost (e.g., value of travel time), planners may need to manage the number of empty AVs cruising or deliberately taking longer routes (or traveling slower) to align with the next passenger’s pickup time and location. The amount of additional pricing could be exactly the same as the travel cost difference between onboard AVs and zero-occupancy AVs.
What are some potential policy recommendations for city transportation departments looking to mitigate the impact of AVs?
Potential policy responses to AVs should focus on managing the impacts of automobile-based travel rather than regulating AVs as a distinct vehicle class. Pricing is likely the most effective tool for reducing the negative externalities of auto travel, including congestion, VMT, and emissions. If pricing is not politically feasible, cities can employ curb management strategies that recognize the higher productivity of shared and for-hire vehicles by designating curb spaces for robotaxis and human-driven taxis to efficiently pick up and drop off passengers, similar to bus stops.
At the same time, cities should avoid policies that unnecessarily restrict robotaxis from using curb space available to other vehicles. Instead, curb and roadway policies should be designed to improve the efficiency of limited urban space and advance broader goals, such as reducing congestion, improving public transit performance, and improving transportation safety outcomes.
HOT INDUSTRY NEWS & GOSSIP
Delivery dramaaa or just flexin’ on your AV partners? Serve Robotics released its Q2 results and while revenue climbed 9% over the past quarter, the company noted that its order volume from Uber Eats fell, causing it to slash its full year forecast more than 50%, as it contemplates ending the partnership. But then the other (self-driving) shoe dropped, as Uber announced it had sold off its entire stake in Serve, which had been initially incubated in its Postmates subsidiary. While the spat stems from disagreements over merchant integrations and arrival time reliability, I see Uber’s noisy exit as a louder signal to the broader AV ecosystem: Uber’s got integrations and investments with everyone from Waymo to Aurora to Avride to Baidu to Flytrex to Hertz to Lucid to Nuro to (ok, you get it…) and now it’s making sure they see that Uber’s the one who controls the demand and customer relationships.
Marketplace madness: Following last week’s story, we’ve seen more mobility and delivery platforms post impressive Q2s. Instacart’s GTV grew 14% to $10.3 billion, while revenue crested the one billy mark. Grab’s GMV climbed 21% to $6.5B, while revenue was up 22% to $997M, thanks to strong growth in financial services and GrabMart performance. And at Talabat, GMV was up 11% to $2.9B as revenue climbed 16% to $1.1 billion.
Gig fleet finance: Uber is investing into Galgo, a Chilean motorcycle financier. Galgo, which also operates in Mexico and Colombia, targets delivery and ridehail workers, and does about $100M in ARR. Over in India, Yulu just raised $93M, also for delivery-oriented two-wheeler rentals. (At MobilityVC, we’ve backed Leasy, which is focused on cars for gig drivers.)
A fight of amazonian scale… NYC Mayor Mamdani is looking to crack down on FedEx and Amazon’s subcontracting practices, as he unveiled the Delivery Protection Act. The bill goes after the DSP model, instead pushing the companies with the logo on the box to employee couriers directly, with the goal of improving pay and working conditions.
Birds of a feather: It’s been another week of tough news for Flock, as backlash against the camera network now seems to be a bipartisan effort. Word is also out that Flock tried to add hundreds of thousands of ridehail cameras to its network, before the partnership unravelled. In response to mounting criticisms and abuses, the company just unveiled improved misuse detection and recommendations for lowered data retention.
Billion dollar bike path: Plans to close an eight mile gap in the LA River bike path have hit a snag, as the budget has ballooned to over $1 billion. That works out to over $24,000 per foot, which is less than Spain and France have paid for actual subway projects (yes, like f’ing underground heavy rail metro systems.) While this project was always a tad bloated — with $365 million earmarked in 2016’s Measure M — the path has now suffered from round after round of community input and fancification. Ask people what they want, without subjecting them to tradeoffs, and they’ll always select a gold Ferrari. Maybe it’s time to adopt the alternate plan, which might require partial closures on approximately 20 days per year, but is far better than no path at all.
Micromobility, macro impact! NABSA just released the 2025 Shared Micromobility State of the Industry Report, finding that the world of shared scooters and bikes is in rather healthy stasis. Both trip and vehicle count climbed, despite a small decrease in the number of cities with systems, while 21% of trips were used to connect to transit.
One (Fewer) Car Challenge: LA Metro just unveiled its One Car Challenge, paying people to use one less car. This builds off a previous, successful pilot the agency did with Santa Monica, but has been expanded to include paying people to go from one to zero cars. 2,000 households could earn up to $600 over five weeks. This sounds fun: pick me, pick me!
Modern Delivery, summer break quick hits: Get your own damn bob chorba, Just Eat Takeaway.com is leaving Bulgaria. DoorDash Flavor Fest heads to Atlanta, Austin, Dallas, Houston, Miami, Tampa. Instacart branches into apparel delivery with Academy Sports + Outdoors partnership, ends item price testing. Behind Grubhub’s “Eat the Fees” campaign. USPS looks to AI-powered optimizations. Yum China finalizes Pizza Hut buyout. McDonald’s profit climbs 5%. JET launches voice-powered ordering, Delivery Hero does too. Deliveroo partners with LookFantastic. Cargo trike maker Iceni Cycles folds. Grubhub settlement finally paid out.
A few good links: Orthodox Jews take to scooters (not kosher, says I.) Car insurance payouts fall. (H/T Steve Greenfield.) Zipline launches Rx delivery for Cleveland Clinic. Imagry and UNVI partner on autonomous buses. IL Gov signs ridehail unionization law. NYC considers illegal e-bike buyback program. How Zoox decided its robotaxis were safe. Dallas mulls trolley extension. Transit bus orders climb. Commuter rail systems grapple with rising insurance requirements. Marti expands ridehailing to 10 new cities. Ford expands Lincoln production in U.S. Teamsters challenge CA DMV’s AV trucking regs. Bolt launches ChatGPT ridehail integration. Lyft activates Disneyland partnership. Grab launches AI-powered TNC ordering call centers. More evidence that land value taxes are good. Archer buys Boeing’s flying-taxi and drone subsidiaries. More premium rides, higher platform fees power Uber + Lyft earnings growth. Via’s Q2 revenue climbs 27% to $136M. Uber and Hinomaru Kotsu partner on Tokyo robotaxi deployment.
See you next week!
- Jonah Bliss & The Curbivore Crew