June 18, 2024

Louis I Vuitton

Savvy Car Technicians

The future of rideshare is a remote control rental car

In rideshare, the single largest cost per trip is the driver. It turns out the vast majority of trips have a second driver in every car: a rider with a license.

Imagine a future in which a rental car is as convenient as an Uber at half the price. More specifically, imagine an on-demand, door-to-door rental car (let’s call it an ODDcar for short).

Here’s how it would work:

  1. You hail a car just as you would an Uber or a Lyft
  2. Within minutes, an empty car pulls up to you curbside
  3. Instead of getting into the backseat, you take the wheel
  4. You drive yourself directly to your destination
  5. The empty car departs

How do you get the empty car to and from the rider? Full autonomy will of course answer that question but in a decade or more. In the meantime, we’ll need to rely on remote control — otherwise known as teleoperations.

The autonomous vehicle industry has developed teleoperations capabilities for years, but they have largely been used as a fail-safe for and a servant to the autonomy stack. If we promote teleoperations to be the primary means of vehicle operation, then we could chart a faster, safer, cheaper, more acceptable route to unpersoned vehicles on public roads.

In this article, I will make the case for a consumer offering and one potential technological path to unlock it. Before we dive into the tech, let’s talk about why the ODDcar matters.

Disclaimer: As an Uber alumnus, I own shares of Uber, but there is no company or technology that I’m peddling here. The ODDcar is no lean startup. It requires no less than bringing a new class of vehicle onto public roads, a feat beyond my capacity to orchestrate on my own. By presenting the case here, I hope to “open source” the idea and pose it as a challenge to the industry.

Capturing the value of autonomy before autonomy

The promise of autonomy to the rideshare industry has always been the elimination of the driver, which lowers costs, which lowers price, which ultimately transforms the marketplace. What if we could capture the majority of the benefit of autonomy in a few years instead of decades?

We can — if we design for better utilization.

In major metro US markets, a vehicle active on a rideshare network will typically spend over half its time on-trip. It will spend 30-40% of its time without a passenger, with roughly 10% spent driving to pick up a rider. Of the time without a passenger, we don’t know how much time drivers spend idle vs. proactively driving in the hopes of reaching more fertile grounds, but let’s make some reasonable assumptions. Let’s assume that in addition to the 10% spent driving to pick up a rider, a driver spends another 10% proactively driving around for a total of 20% time spent repositioning.

In traditional rideshare, you have a paid driver engaged with a single vehicle 100% of the time. While it’s true the rideshare driver is only directly paid for a portion of that time, they bear the opportunity cost of that entire time. Their rideshare earnings ultimately need to adequately compensate them for their entire time, since that is time they could otherwise use to earn money at a different job.

With an ODDcar, the paid driver is remote and is only engaged when the vehicle is actively repositioning between trips. Critically, when the car is on-trip, the rider assumes the driving duties, obviating the need for an in-vehicle, paid driver. If we assume 20% of a car’s time is spent repositioning between trips, then one remote driver can support 5X the number of vehicles relative to traditional rideshare drivers.

There will of course have to be measures in place to ensure remote drivers are responsible. I envision all remote driving operations to be performed in dedicated facilities under tightly controlled conditions, where sobriety can be verified and driver performance monitored carefully. Also, there would need to be technologies that would securely “bind” an operator to a vehicle while in operation, creating a clear record of custody and accountability. I imagine this will be written into legislation, but even without, the ODDcar fleet operator will have its own strong incentive to do all of this effectively.

A real unit cost advantage

What does this ODDcar model mean in dollars and cents? Let’s make up some numbers!

To get a car out on the road for a day, real costs are incurred. Some of these costs have traditionally been paid by the rideshare network while others have been paid by the rideshare driver. For this exercise, we don’t care. If somebody bears the cost, it goes into this list. Major costs fall into one of the five categories below. Let’s assume each category roughly accounts for the following (play along and make your own assumptions!):

Now let’s say the cost of getting a rideshare car onto the network for an hour is $25. At that rate, the costs per hour break down as follows:

(I want to stress that all of the numbers here are completely made up. Actual numbers vary widely depending on who you talk to, which market you’re in, which driver segment you observe, or whether Jupiter aligns with Mars. Rather than get bogged down in particulars, I present these toy numbers to illustrate a larger point.)

Finally, let’s wave our magical hands and see how these rideshare costs stack up against those of an ODDcar and a fully autonomous Robotaxi.

The most important line is the first, where driver net earnings drop from $17.50 to $3.50 to $0.

The leap from $17.50 to $3.50 comes from the 5X leverage ratio I assumed in the discussion of utilization. Only 20% of a paid driver is necessary to get an ODDcar onto the network for an hour.

The maintenance and depreciation costs increase from left to right as you capitalize more specialized hardware, sensors, and compute into the vehicle itself. The increase from the left column to the middle basically amounts to an incremental but not multiplicative bump on the cost of the vehicle. Again, make your own assumptions. The increase from middle to right recognizes that the full autonomous stack is going to be expensive for quite a while.

For the ODDcar only, I assume an extra $1.25 in “other” for video transmission over 4G to support teleoperations, again assuming that teleoperations are only needed during the 20% of the hour spent repositioning. 5G will further reduce costs and improve performance.

Regarding insurance, I assume that the ODDcar operator would provide insurance to riders and remote drivers alike. It is helpful to consider on-trip vs. off-trip cases separately, as the risk profiles and insurance rates for each state would be quite different:

  • On-trip, ODDcar insurance would fall under a car rental classification. Relative to rideshare, car rentals have far lower minimum coverage mandates and fewer occupants to insure, so there is reason to believe there are significant cost savings here.
  • Off-trip, I’d expect remote operation to add incremental risk relative to an in-person driver. The cost of that risk is mitigated due to two factors: a) remote driving is used only for only a small fraction of every hour and every trip, and b) remote driving is only performed at low speeds in low risk environments with no passengers.

On balance, ODDcars promise an insurance cost savings relative to rideshare, but I’ll defer a deeper discussion for another time. For this exercise, I’ve conservatively assumed costs would be constant to demonstrate that you don’t have to believe insurance will drop in order to believe there is an opportunity. Moving to the right, I’ve generously granted the Robotaxi insurance savings for surpassing human safety.

Finally, I’ve held fuel constant across all categories (electrification would only further benefit the latter two columns).

The grand takeaway is this: As you move from Rideshare to ODDcars, unit costs decrease by a whopping 48%. As you move from ODDcars to Robotaxis, you get an incremental 12% decrease. A full 80% of the economic step change promised by full autonomy is captured by the ODDcar.

A dominant category in the market for rides

How are ODDcars positioned in the rideshare ecosystem? Put simply, for riders, an ODDcar is as fast as an UberX, cheaper than an Uber Pool. ODDcars take riders directly from origin to destination at human driver speed with no detours.

In exchange for a fast ride and a market leading price, the rider bears the burden of driving. The extent to which driving oneself is a deterrent is open for debate. However, we know that in mass market travel and transit, price and schedule are always king and queen. Further, consumers will endure endless humiliations in order to optimize for those two things. When you consider that driving yourself can actually be seen as a benefit for many riders — those who fear unsavory interactions with drivers, for example — you have the makings of a winning product offering.

Further, an ODDcar’s cost advantage does not scale linearly with the length of the trip. The longer the trip, the more pronounced the cost savings. ODDcars can dominate rideshare trips of 20-60 minutes (think airport trips, commutes) and unlock longer trips that are currently too expensive to service (multi-destination, errands, day trips, inter-city travel).

How big does this get? As a true step-change cost innovation, I believe it will be the next growth driver for the rideshare market. It will grow the market in the way that UberX eventually dwarfed its Uber Black predecessor. In other words, I believe this is huge.

Transit in a pandemic

In a time of global pandemic, it’s better to separate riders and drivers by hundreds of miles than by two feet.

More seriously, the COVID-19 crisis has turned ridesharing into a potentially powerful vector for community spread. By minimizing the number of people sharing a vehicle’s airspace at one time, ODDcars can be a critical piece of a more crisis-resilient transit landscape. ODDcars eliminate human-to-human contact between riders and drivers and between pooled riders. ODDcars will not solve all of the public health risks associated with ridesharing today (between-trip vehicle sanitization, for example), but I will defer discussion of that for another day.

Getting to safety

What is the state of teleoperations technologies? When networks are strong, teleoperations rival in-person performance. Building the teleoperations safety case is focused on the inherent vulnerabilities of networking (latency and reliability).

Issues related to latency increase dramatically with speed, so teleoperation is more naturally suited for slower speed applications. To bring the ODDcar to life, even the modest ability to travel up to 25mph on local suburban roads is sufficient to address a sizable market. What technologies do we need to unlock that domain?

My working hypothesis is that a human-in-the-loop 100% of the time plus a reduced set of achievable autonomous capabilities plus a reduced autonomous hardware stack may be the fastest path to unpersoned vehicles on public roads. Think about it this way: If a human is always in the loop for all of the object recognition, long range planning, and high level decision making, then the biggest safety gaps to fill are all short range, low latency emergency maneuvers: emergency stopping, forward obstacle detection, rear object detection, graduated stopping, lane keeping, and parallel parking. Incidentally, these sound a lot like ADAS capabilities that have been productionized for years. That’s not to say that the work is done. There is plenty left to do. The question is simply whether this path is the fastest path.

Will this be safe enough? Will it be acceptable? Nobody can say for sure, because there is no clear definition of what “safe enough” is. Even those regulations that have been written will surely be scrutinized as unpersoned vehicles approach production readiness. That said, imagine in your mind the first unpersoned vehicle allowed to operate on public roads at scale. You’d probably picture a vehicle with capped speeds, no passengers, and a dedicated human in the loop at all times during unpersoned operation, all features of the technical approach I’m describing. The ODDcar rental model gives us a way to get real economic leverage from even this very limited set of capabilities and turn it into a transformative (and profitable) consumer offering.

Some prominent voices argue that teleoperation is unsafe at any speed on any public road unless there is so much autonomy that you would obviate the need for teleoperation in the first place. Other prominent voices argue just the opposite. One industry veteran said to me, “There are those in the AV industry who aren’t serious about launching products in the real world, and then there are those that believe in teleop.” As is the case with any frontier tech, there are a range of opinions about which paths might bear fruit.

Autonomy’s uncanny valley

Robotaxis can feel magical, but for anyone who’s tried one, they can also be terrifying. They are jerky, unintuitively cautious, and just plain eerie. There will likely be a significant period of time between when the technology is safe enough for public roads and when the experience is not deeply unnerving. Consider also that the first vehicles to be certified as safe for public roads may have severe limitations on speed and service area, making for the world’s most infuriating taxi service. This is the uncanny valley of autonomy.

Contrastingly, an ODDcar only invokes its driverless technology when empty and repositioning. Under those conditions, it can optimize routing and speed for safety without regard for passenger safety, comfort, or convenience. As soon as the rider enters the vehicle, it is as safe, as familiar, as capable, and as fast as a regular car. The ODDcar is a wonderfully non-threatening way to acclimate the general public to unpersoned vehicles roaming their streets.

Note that the uncanny valley principle would suggest that an ODDcar rental car model is the right MVP for any driverless tech, whether it be powered by autonomy or teleoperation.

Two points of leverage

The ODDcar offers two important points of leverage. The first is the obvious one I’ve been discussing: One paid remote driver can support more vehicles per hour.

The second is less obvious but perhaps just as critical: For a given set of driverless capabilities, you can service trips between serviceable areas, not just within.

Imagine you have developed an unpersoned vehicle that is safe for public roads, but it can only be used on local roads at speeds of no more than 25mph in selected pre-mapped areas. That would only allow you to run a very limited Robotaxi service. However, if you used that technology for repositioning an ODDcar, then a trip need only start and end within one of these pre-mapped areas. The bulk of an ODDcar trip could be over terrain that would be treacherous to your driverless stack, but since the rider is at the wheel, it would perform like a regular car. It would be as safe as a regular car, and critically, it would be as fast as a regular car.

Put it this way: operating a shuttle service within the San Francisco financial district is not an interesting business. Operating a shuttle service within Stanford campus is not an interesting business. Operating a service that can support trips between the financial district and Stanford begins to smell like an interesting, scalable business.

Note again that this idea would argue that an ODDcar rental car model is the right MVP for any driverless tech, whether it be powered by autonomy or teleoperation.

Better jobs for drivers

Being a remote driver can be a better job than being a rideshare driver today. Remote drivers will likely be directly employed and will be offered both the protections and benefits of employment. Earnings will be more predictable, and remote drivers don’t bear the costs and uncertainty of owning and maintaining a vehicle.

Further, remote driving is safer and more comfortable. There is no risk of bodily harm, either from traffic or unsavory passenger interactions. Remote drivers are not at risk of theft or vandalism. If you’ve ever driven for a rideshare company, you know the hell on earth bathroom breaks can be. And of course, highly communicable airborne diseases cannot be transmitted over 4G.

You may ask, doesn’t this model enable companies to outsource remote driving jobs to low wage countries? For the foreseeable future, there are several natural forces that would keep the jobs domestic and likely intraregional. First, having a domestic license and familiarity with local driving customs will be important factors in delivering a trustworthy service. Second, telecom speed matters for teleoperations. There are still significant latency advantages to being physically proximate. Third, because ODDcars get so much more leverage out of paid drivers, the potential savings from offshoring just matter much less as a percentage of overall costs.

Unleashing potential energy

Truly disruptive technologies release economic potential energy. AirBnB unlocked unused real estate. Uber unlocked underutilized vehicles and drivers. ODDcar unlocks a huge store of potential energy: a rider’s ability to drive themselves, a valuable skill that lies fallow in the backseats of millions of rideshare rides every day. That is unutilized economic capacity that doesn’t exist in other applications of autonomy: food delivery, freight, micrologistics, scooter repositioning, fork lifts, or yard vehicles. It’s unique to passenger transit, and it’s why I believe passenger transit is going to be the first huge category for driverless tech.

ODDcars matter regardless of underlying technology

In this article, I’ve attempted to make two points: 1) I argue that the ODDcar car rental model will be the first application of driverless tech at scale in the rideshare industry, and long before Robotaxis take over, it will become the dominant model in the rideshare industry. 2) I argue that leveraging teleoperations as a foundation for driverless technology can help to get an unpersoned vehicle onto public roads faster, safer, and cheaper.

These points taken together describe a disruptive opportunity and its technological roadmap, but they can be considered independently. Of these two points, by far the more important is the first. You can power ODDcars with fully autonomous solutions, and it still will be a more compelling product offering than Robotaxis for a very long time. I’d contend that even without teleoperations, unless autonomous technologies hit the market all at once with comprehensive capabilities that can be deployed broadly across markets, the ODDcar rental model is the inevitable first phase of deployment in the rideshare industry. Far from a niche stopgap, it will grow to upend and dominate the rideshare industry long before the world is ready for a Robotaxi at scale.

OK, what’s next?

The rideshare industry has always assumed that the big economic step change will only come once we can get the car to drive itself. I believe the step change comes when you can get the paid driver out of the car.

The idea that the ODDcar will gain hold long before Robotaxis is more than a prediction. If true, then it compels us to act. If the past five years have taught us anything, it is that autonomy is not going to be a big bang event. It will roll out in creeping increments over a long period of time, and each increment will be hard won. How effectively we roll out depends greatly on what use cases we deliberately target. In that light, are we investing in the right technological paths? Are we chasing the right regulatory frameworks? Are we targeting the right operational design domains? Are we building the right vehicles?

My hope is that putting these ideas out to the universe will start a new discussion. If there are valuable ideas here that withstand the scrutiny of peer review, then maybe we can begin to hone a new vision for the future of driverless technologies.

Special thanks to Zoubin Ghahramani, Vijay Subramanian, Evan Arnold, John Butler, Yvonne Hung, and Michael Kim for reviewing drafts of this article.

Francis Kim is a product strategist, AI enthusiast, and angel investor and an alum of Uber, Index, Limewire, and ProfitLogic. Follow him at medium.com/@franciskim or on Twitter @franciskim.