Why Tesla Will Win The Autonomous Driving Race

Autonomous driving technology has become all the rage in the last year — 2016 saw new companies announced and acquired at high valuations (Otto , Cruise) and bold claims made (Tesla, Volvo). Significant resources are pouring into the field and almost every major player in the automotive world is now involved (link); suffice to say this will be something to watch for in the next 3–5 years. So, who might get there first?

This is an interesting question. It appears that industry-wide, companies are following one of two general approaches to achieving fully autonomous driving. I’m going to generalize them as follows:

  1. Incrementally reach full autonomy. This approach is similar to cloud services — new features don’t need to be “packaged” into discrete products, they can be released piece-meal as they’re ready for customers to use.
  2. Target full autonomy with no intermediate steps. “Solve” autonomous driving and release products that are fully autonomous (Level 5), wherein a user does not need to do anything other than disclose a final destination.

Let’s consider approach #2 first. This is the approach of companies such as Waymo. This approach can work well in sandboxed use cases such as limited-area taxi services, etc. If fully autonomous driving is analogous to general AI, then this use case would be narrow AI. Pursuing full autonomy broadly (i.e. not just in sandboxed use cases) with no intermediate steps is a lofty and ambitious goal — as a society we certainly need such goals and I commend companies like to Waymo for pursuing these. But they’re not practical. Here are a few reasons why:

  1. It’s all or nothing. Waymo has gone so far as to not even bother with a steering wheel. The lack of a steering wheel implies that such products will not be production ready unless every single safety-relevant edge case has been tackled. In software terms, you can’t ship with any bugs — almost every bug would be a launch blocker because of safety concerns. We can’t apply the 80/20 rule here — all known issues need to be fixed to enable launch.
  2. The Unknown Unknowns. Even if we were to fix every known bug, that would only cover the set of known knowns. What about the set of unknown unknowns? Approach #2 assumes that 1) all possible problems are already known or knowable before launch, or 2) if not known at launch can be identified and addressed before they lead to large scale safety risk. I say large scale because if this approach leads to a vastly improved user and societal experience overall, then societally we may be willing to accept some level of safety risk.
  3. Scaling. Waymo is currently testing its cars in four cities. But four cities is a very, very limited data set. The internet was abuzz a few weeks ago about how Waymo had way fewer instances of the driver having to take control in 2016 (what are called disengagement events), but that better be the case — they’ve been driving around in the same cities for years! The real challenge for Waymo and full-autonomy-only proponents is that of scaling. How do you take the learnings from a few data points and ensure that you’re covering ALL possible scenarios? Waymo says that they can take real-world data and simulate related events. This is fascinating and creative for sure, but still underestimates the number of possible edge cases.

It is in the above scaling sense that approach #1 has a huge advantage over approach #2. Even if, with the most high-precision maps ever and the best sensing technology ever, we solve the problem of driving on the highway or on known city streets without problems, what do we do about altered traffic patterns, or less travelled roads in the countryside? The short-term route alteration problem is much more challenging than it seems. The challenging question isn’t how companies will update this data to reflect the new reality, but how they will approach autonomous driving when the new data doesn’t exist yet. Approach #2 is very analogous to achieving general AI, which is further out in the future and more nebulous than it seems.

Now let’s consider approach #1 .Tesla is the poster-child of this approach — over the last 30 months, they’ve added incremental self-driving features with a certain degree of success. One of the key philosophical considerations here is the engineering versus customer angle. From an engineering standpoint, the prospect of solving autonomous driving is an enticing one. Udacity’s Nano Degree has received a lot of attention — and rightly so. Engineers want to build new technology and we have a bias for perfection. But customers want improvements to their quality of life — and they’re generally willing to pay for it. So there is a real tradeoff here — do we wait for X years in the future to improve customers’ lives a lot or do we give customers some improvement now and keep iterating gradually until we get to the same end goal in possibly the same amount of time. I think this one is a no brainer.

What Tesla Has Gotten Right…

Tesla’s genius breakthrough isn’t in the machine learning itself (other companies might have even better tech — who knows?). Its genius is in the program structure and business model it’s put in place. Machine learning is heavily dependent on data and Tesla is the Google of the autonomous driving industry. Their — increasing — lead with regards to data is their moat. To match Tesla’s fleet of say 100,000 units on the road, assuming a cost per car of $25,000 — you’d need $2,500,000,000 in initial capital. Add to that 100,000 drivers driving say, 20 hours a week at $15 per hour, the minimum annual cost of operation is about $1,500,000,000, excluding the cost of fuel. Assume that it takes such a fleet 5 years to reach 90% autonomy — that’s $10 billion in up-front investment, to match Tesla’s capabilities today.

There are a couple of ways for competitors to approach this: they could start a consortium to pool resources and compete with Tesla. But this won’t work because a committee isn’t going to be a single-minded competitor. They could look to an OS provider (think the Android of autonomous driving) who could aggregate data at scale and leverage it to leapfrog Tesla. This could work but is challenging because a whole lot of companies are going to have to sell (or subsidize) expensive hardware. Sure that seems doable when you’re competing with 100,000 Teslas. But what about when it’s a million? Also, I don’t think traditional companies will have the appetite to mix into lower margin products to such an extent, and I am doubtful customers will be willing to foot the bill on this one. This option would pose a big coordination problem and also looks unlikely.

Eventually, the battle for autonomous driving will be won on the roads less travelled, literally and figuratively. It is relatively trivial to automate driving on the interstate or other highways, but where I think Tesla will pull ahead of the pack decisively is on smaller roads, where the structure and layout of roads, stops, merging lanes, etc. is generally more varied across regions. For example, I remember from my time in New England that roundabouts are a lot more popular there than in many other parts of the country. In any situation where there is deviation from the norm, because Tesla is able to capture real-world driving behavior data quickly, it will be much more likely than any other player to have the training data required to automate driving in that condition. The data advantage will only grow with the launch of the Model 3. This massive data set allows Tesla to continuously improve its algorithms. Every new Tesla benefits from the rest of the network, and the network benefits from the addition of every new Tesla. This is a virtuous cycle — described well by Prof. Scott Golloway of NYU, here.

Initially, the system will be best suited to roads that provide the most amount of data (interstates, other highly trafficked highways) or have less variance in features (again interstates, major highways) and it’ll be relatively poor on roads less travelled simply because the company won’t have enough (or maybe even zero) training data. It’ll be years before Tesla has any meaningful amount of data for the long tail of roads and routes in the U.S — let alone the whole world — so don’t expect the steering wheel to go away any time soon. However, I don’t think the lack of a steering wheel is the holy grail here; if customers don’t need to be actively engaged in paying attention to the road for say, 80-90% of their total miles driven, that would be a huge deal.

And Tesla will accomplish this in production first.

DISCLAIMER: I am long TSLA and own stock in the company.