Patrick McGee of the Financial Times wrote that “there is a growing recognition that getting the self-driving algorithms right is merely an entry ticket for the much bigger challenge of commercialization,” adding that “the technical know-how to manage a fleet that competes with the likes of Uber and Lyft on timely pickups” is critical to commercialization.
Apples And Oranges
Competing with Uber and Lyft on timely pickups represents an enormous challenge but also brings up an “apples and oranges” comparison. Peer-to-peer ride-hailing services are “convenient” for travelers because the businesses flood city centers with vehicles so that a ride is always nearby. This convenience is affordable for Uber and Lyft because they only pay drivers when a passenger is on board, and they don’t incur any costs from the empty vehicles.
In fact, the convenience that ride-hailing businesses offer comes at a cost of efficiency and congestion. According to a report from Schaller Consulting, these vehicles add 2.8 miles for every mile of personal driving they eliminate. As a result, peer-to-peer ride-hailing services have made traffic 180% worse in cities.
For autonomous service providers or a service provider that owns its conventional vehicles, every mile and minute incurs a cost. Providing a convenient service will have to be balanced with vehicle utilization targets, or paid miles. Offering “timely pickups” while also ensuring a sustainable level of paid usage is really difficult.
10,000 Times Harder
Waymo’s former head of business development, Shaun Stewart, told McGee that getting self-driving right is “just one milestone. How do you succeed in the commercial endeavors around the technology? It’s a completely different challenge and requires a completely different skill set and experience set.”
Cruise president Dan Ammann told McGee that it’s “probably 10,000 times harder” to deliver a commercial autonomous service at scale than to demonstrate that the vehicle can drive down the street.
Deciding in real time which vehicle should go where may seem simple when thinking of a single ride request from point A to point B. But to do this at scale requires evaluating an enormous number of variables that grows exponentially with the size of the operation.
Imagine fielding hundreds or thousands of ride requests simultaneously and instantly assigning hundreds of vehicles to riders with predictable ride times and wait times, and with predictable revenue per mile traveled. Among the variables that need to be considered are current and future traffic conditions, vehicle locations, capacities and available energy, and even possible failures. Vehicles may need to be reassigned midroute to accommodate incoming requests.
For an autonomous vehicle business to be successful, these parameters — vehicle utilization and passenger convenience — will need to be manageable and controllable. Optimization algorithms will need to go beyond the current state-of-the-art models in operations research and artificial intelligence.
“The self-driving system is a core competency, it’s necessary — but it’s not sufficient to create a new business,” James Farley, the president of new business at Ford, told the Financial Times. Transforming autonomous vehicles into a successful business requires more than safe driving; it requires fleet optimization at levels few AV tech developers have contemplated.
A first step to optimizing fleets of autonomous vehicles is to think of shared transit as the primary use case and business case for AV fleets. Traffic gurus at MIT developed an algorithm and found that 2,000 10-person vehicles could handle 95% of New York’s 14,000 cabs.
Adopting a shared model for most or all AVs should drive vehicle design (larger, multipassenger vehicles) and feature offers like Wi-Fi and entertainment to make slightly longer journeys more attractive than driving due to the opportunity to get things done during the trip.
Even without algorithmic optimization, mobility services could be more efficient incentivizing and rewarding the smart routing of vehicles (forcing them to take specific routes) and altered travel times to get people to book rides at off-peak times, according to Carnegie Mellon engineering professor Sean Qian. He studied today’s ride-hailing data, but there is no reason these kinds of rules could not be baked into AV booking apps, too.
Asking The Wrong Questions
Comparing autonomous fleets to today’s peer-to-peer ride-hailing business — thinking that autonomy simply involves removing a human driver — is an overly simplified view. If operator-owned robotaxis replaced every Uber and Lyft vehicle, with the supply-demand mismatch they have today, the companies would likely lose even more money. The excess vehicles would still be empty most of the time, and those empty vehicles would become very expensive — someone has to absorb that cost.
Earlier this year, Waymo announced it would not build autonomous cars, telling The Telegraph (via Observer) that it would instead focus on creating “the world’s most experienced driver.” Many of us are anxiously awaiting AVs to be safe enough for prime time, but that still doesn’t guarantee that the vehicles will be deployable in fleets that actually deliver a better service than their predecessors.
I believe what cities really need is services that move more people with fewer vehicles, reducing traffic and emissions. It could be argued that with urban congestion and pollution becoming worse and deadlier, the world needs the business solution even more than the self-driving solution.