We are entering the Slope of Enlightenment for self-driving cars.

Since Waymo’s founding 15 years ago, billions have been invested into self-driving cars with little revenue to show for. While the hype got well ahead of the technology, engineering teams have been making steady progress in solving one of the hardest problems in robotics.

After years of disillusionment, we’ve recently seen a new cycle of announcements and progress:

• Waymo is slowly expanding to more cities. After launching in Scottsdale in 2020, it took them another 4 years to open San Francisco. Further expansion in the Bay Area and launching Los Angeles is on the horizon.
• Tesla has been teasing their big robotaxi announcement for months but just pushed it back from August to October. In typical Elon fashion, Tesla will probably get there, just not on his timeline.
• GM just announced they are restarting Cruise. Zoox is launching in Las Vegas and various Chinese players have made progress both in China and in the US. Softbank is trying to create a new autonomy alliance of legacy car makers.

‍

Diverging technical approaches

There has been a lively debate about the diverging technical approaches. Can full autonomy be achieved with cameras alone (Tesla’s approach) or is a multi-sensor stack needed combining cameras, lidar and radar (Waymo’s approach)? Tesla has also been talking about moving to end-to-end machine learning and replacing large parts of their deterministic, rules-based software with neural net decision-making. Waymo might be doing the same but not talking about it publicly.

Let’s put the technical considerations aside for a minute and focus on the consumer perspective. At the end of the day, consumers will make choices that drives adoption, direct hundreds of billions dollar of spend and crown the winners in this market. Americans drive 3.2 trillion miles a year and spend $500B+ on new cars each year.

‍

Experiencing autonomy

While the industry has a gradual definition of autonomous driving, once you experience autonomy driving around San Francisco in a Waymo, you notice the difference between a driver assist feature like Tesla’s FSD or MobileEye’s products and cars that drive actually themselves.

Levels of autonomy shown on a continuous spectrum:

‍

What users actual experience:

‍

The self-driving chasm between L3 and L4 is a huge jump, both in terms of technical difficulty and in the resulting user experience. Actual autonomy starts at L4 when the car takes full responsibility (and liability) for a meaningful time of users’ driving needs. It can drive my kids to soccer practice or you take a nap while cruising on the highway. Only when the user can use the time for something other than driving (sleeping, reading, working, watching movies etc) will we experience a transformational new product.

‍

Will we still own cars?

Some believe robotaxis will displace car ownership. You ditch your car and just pay per mile. This is unlikely to happen for a few reasons:

• Robotaxis are expensive machines that require high utilization to be economically viable. In contrast to Uber drivers that provide a flexible supply of rides when demand is high, robotaxis that are only earning revenue a few hours a day will be too expensive to operate. As hardware costs come down and Telsa launches their robotaxi service that deploys privately-owned cars, prices per mile will decline in high liquidity markets like dense urban areas taking more share from Uber and co but not replacing car ownership in a massive way.
• In its early years, ride hailing companies like Uber and Lyft made the same argument that we can get rid of our cars. It didn’t happen. Customers like the flexibility of having a car on their doorstep that is available without delay and can take them anywhere they want. Usage patterns are also a major factor: most people need their cars at the same time. So while cars are an underutilized asset, they are underutilized at the same time when people don’t need a car (ie at night when people are sleeping or during the day when people are working).
• Customers care deeply about their cars - the color, the interior, the shape, size etc. Cars become an expression of people’s identity with funky stickers and specialized license plates. Customers want a choice and are willing to spend a lot of money buying what they enjoy driving. 

Zooming out, ride hailing services like Uber and Lyft make up less than 3% of miles driven in dense urban areas. Robotaxis will not change that in a meaningful way for the reasons outlined above. The vastly bigger market for self-driving technology will be selling it to car owners. 

‍

Liability and safety records

Once liability of driving transitions away from human drivers to the companies that sell autonomous systems, safety records will play a crucial role in buyer decision making. Studies like these from SwissRe will get formalized into detailed safety records (measured in fatal and non-fatal accidents per million miles) and regulators will step in to mandate safety measures. Systems that are deemed unsafe will get banned by regulators or will fail due to a lack of consumer trust. Safety will converge on a high level which will be safer than human drivers. We’ll take safety for granted in the same way we do when we sit down in a commercial airplane.

‍

Winning business models

If we follow this consumer-driven perspective, broad adoption of self-driving technology will come from car manufacturers integrating it into their vehicles. It is unlikely that most OEMs can stem the technical challenges themselves - so far the problem has proven too hard to solve. More likely is that the leaders in autonomy will license the technology to the OEMs. Initially the technology will be expensive and only sold in high-end cars. Over time prices will come down and adoption will expand. 

If autonomy is rolled out geographically (as Waymo has been doing it), new revenue models might emerge where OEMs sell the hardware itself as part of the car (”self-driving ready”) and then customers can purchase autonomy packages by region. If you go on a trip to another city, you could add that package for a weekend. Or you pay for a roaming plan similar to cell phones. If Tesla’s approach works for L4/L5 (cheaper hardware stack and end-to-end machine learning that is not geographically limited), their incentive will be to sell as many cars with the technology as possible. If the technology is as transformational as I think it will be, other OEMs will be pressured to license the technology to not lose market share.

Integrating self-driving technology in series production cars will be a huge undertaking. The sensors, compute hardware and required power for autonomy will mean major changes to a car’s design, resulting in long integration and development timelines. If Waymo’s technology were ready to be licensed to an OEM today (which it’s not), it would likely take 5+ years for those cars to be available to the public. Until then, robotaxis are the best way to monetize and mature the technology.