Self-driving cars operate using machine learning. Machine learning algorithms have become commoditized, even open source. Data, not algorithms, is what confers competitive advantage. The company with the most data wins. Currently, Tesla (NASDAQ:TSLA) has access to vastly more driving data than any other company.
The Alphabet (NASDAQ:GOOG) (NASDAQ:GOOGL) subsidiary Waymo, once widely recognized as the leader in self driving, had a cumulative 2 million miles of driving data in October 2016. Tesla's cars with "full self-driving hardware" (HW2) are currently driving over 1 million miles per day. (The math: Tesla sold at least 14,000 HW2 cars in Q4 2016. Assuming it has sold 1500 cars every week of production - the average for 2016 - since the beginning of the year, there are 27,500 HW2 cars driving an average of 37 miles per day.) As best as I can surmise from public information, Waymo's current fleet of test vehicles numbers in the hundreds - less than Tesla adds to its fleet every week - with no indication yet of concrete plans for significant growth. Waymo should be worried. Tesla's data advantage will continue to widen until Waymo or another competitor adds tens of thousands of cars to the road.
Only a fraction of the overall miles are driven with Tesla's Autopilot driving assistance software engaged. However, even if just 5% of overall miles are Autopilot miles, at its current rate Tesla would still surpass Waymo's cumulative all-time 2 million miles every six weeks - and that rate accelerates every week as new cars are added to the road. As Enhanced Autopilot and then the first features of Tesla's Full Self-Driving software are incrementally rolled out to customers, the software will be able to be engaged in more driving situations and so the percent of overall miles driven with software engaged will likely rise. Moreover, Tesla has described plans to run its Full Self-Driving software in "shadow mode," so that it can test its software against human drivers when the software is not engaged.
Two arguments that Tesla is behind Waymo in self-driving are: 1) Tesla's test cars have more disengagements per mile than Waymo's and 2) Tesla's HW2 cars don't have LIDAR, which is essential for self-driving. There is overwhelming reason to doubt that the disengagements data provides us with any useful information about Tesla's self-driving performance. There is also a good case that Tesla can achieve full self-driving without LIDAR.
Photo credit: Norsk Elbilforening.
Disengagements data isn't useful
A disengagement occurs when a driver turns off a car's self-driving software and takes control of the vehicle. Paulo Santos points to California DMV data from 2016 showing that Tesla test cars were disengaged 1,709 times more frequently on California's public roads than Waymo test cars. On the face of it, this would seem to indicate Waymo's self-driving performance is 1,709 times better than Tesla's. A closer look, however, should lead one to doubt this interpretation.
For Waymo, the sample size is 635,868 miles, roughly a third of the total 2 million miles Waymo cars had driven as of October 2016. For Tesla, the sample size is just 550 miles. That's 1,156 times smaller. 550 miles is less than two full battery charges of a base Model X. If we take this data at face value, the most worrying aspect for Tesla should not be the rate of disengagements, but the fact that it appears to be hardly testing self-driving cars at all!
A much more plausible explanation is that Tesla does virtually all of its testing outside of California in states like Texas and Arizona that do not collect data on disengagements and where companies like Waymo and Uber (Private:UBER) are known to be testing self-driving cars. In fact, according to a Tesla spokesperson, the 550 miles that occurred in California were primarily for the purposes of video production, not testing. Since the demands and goals of video production are entirely different from those of testing, the number of disengagements is not indicative of Tesla's self-driving performance.
Even if we assume that in 2016 Tesla's self-driving test cars had much worse performance than Waymo's - certainly a plausible assumption given Waymo's impressive track record - Tesla has access to vastly more data and therefore has the potential to make progress at an immensely faster rate. By November 2017, Tesla's HW2 cars will be driving an order of magnitude more miles every week than Waymo's cars have driven in its entire history stretching back to 2009. The more data a company has, the faster it can train its machine learning algorithms. All else being equal, Tesla's self-driving software will eventually drive laps around Waymo's if it doesn't ramp up its testing dramatically. It's just a matter of time.
Photo credit: Grendelkhan.
LIDAR probably isn't necessary
However, all else isn't equal. The number of miles driven is irrelevant if cars don't have the right sensors for full self-driving capability. This is why Tesla's cars with the old Autopilot hardware (HW1) won't ever achieve full self-driving, despite accumulating hundreds of millions of miles of driving data. Critics like Paulo Santos argue that "Tesla by foregoing the use of LIDAR puts itself at a significant disadvantage." This argument is understandable. Tesla is bucking a trend that almost all companies in the self-driving space are on board with.
Waymo notably relies on LIDAR as its primary sensor. LIDAR is highly accurate in clear weather conditions, but - and this is too often overlooked - it can't see through heavy rain, fog or snow. Its accuracy as an average across all weather conditions, then, is not great. Tesla's HW2 cars use radar (as do Waymo's) that can see through rain, fog and snow, with not much less accuracy in clear weather conditions. Even in HW1 cars, Tesla's radar system will cause Autopilot to apply the brakes when a visually obscured car stops ahead. Given the trade-off, Tesla's choice of sensor suite is reasonable. It makes sense to just use a cheap, compact sensor that works well all the time rather than to insist on supplementing it with an expensive, bulky sensor that works a bit better most of the time but fails some of the time.
The main selling point of LIDAR is its ability to create a detailed 3D map of a car's surroundings by rapidly pinging surfaces in range with laser pulses. Each returned laser pulse forms a point, and a cluster of nearby points forms a point cloud that corresponds to an object. Tesla believes it can forego LIDAR because it has developed software that uses radar data to create a similar detailed 3D map composed of point clouds. The Achilles' heel of radar is reflective metal objects like road signs which can cause a car to brake unnecessarily. Tesla is relying on fleet learning assisted by human drivers to recognize benign metal objects and tag them as safe based on their GPS coordinates.
At least some experts vindicate Tesla's stance on LIDAR. Jeffrey Miller, a self-driving car engineer interviewed by Wired, says that Tesla can "absolutely" achieve full self-driving without using LIDAR. Jianxiong Xiao, CEO of the self-driving car startup AutoX, expressed a more skeptical view, but conceded that Tesla's cars may be able to drive more safely than humans without using LIDAR.
Nonetheless, Tesla is taking a big risk by advertising its current sensor suite - comprising eight cameras, twelve short-range ultrasonic sensors, and one radar - as sufficient for full self-driving. If it turns out that additional sensors are in fact required, Tesla may face a huge customer backlash, a blow to its reputation, and even lawsuits. Tesla has thrown its hat over the wall and committed itself to making the HW2 sensor suite work, at great cost if it fails.
A competitor could still take the lead - if it's bold enough
Tesla's main advantage in the self-driving space is its access to a vast amount of driving data. Its main disadvantage is the comparatively small volume of cars it can produce to collect that data - so far it produces less than 100,000 cars per year, with a planned ramp to 500,000 per year in 2018. A large automaker like General Motors (NYSE:GM), which produces nearly 10 million cars per year, could quickly overtake Tesla in driving data by equipping hundreds of thousands or millions of new cars with full self-driving hardware, perhaps in partnership with Waymo or Intel (NASDAQ:INTC), which recently acquired Mobileye. Here, the incumbents have a powerful advantage, if they are bold enough to seize it.
General Motors appears to be the most forward-looking of the incumbent automakers when it comes to self-driving, although perhaps it is simply the most public about its ambitions. But as Arne Alsin observes in Forbes, General Motors is also far more complacent and slow-moving than it could be, doling out dividends rather than investing aggressively in the future of automotive mobility. GM is reportedly planning to roll out "thousands" of self-driving test cars in 2018. It's unclear exactly how many thousands. A test fleet of just a few thousand cars will still leave GM far behind. By January 2018, Tesla will have nearly 100,000 HW2 cars on public roads - not counting the Model 3 - on the conservative assumption that it sells 1500 cars per week. If Tesla meets its target of producing 5000 cars per week by the end of 2017, the hill will get rapidly steeper to climb for any competitor.
There is a limited window of time for incumbents to act decisively before they risk losing their upper hand in production capacity. If Tesla is the first company to achieve full self-driving in commercially available vehicles, it will enjoy a stratospheric boost to its reputation and brand and a dramatic surge in demand for its cars until a competitor can offer an alternative. This demand might allow Tesla to accelerate its production ramp, just as it pulled up its production timeline by two years when it received an unanticipated level of demand for the Model 3. It could potentially expand Gigafactory 1 and accelerate construction of Gigafactory 3, 4, and 5. Gigafactory 1 is expected to be capable of supporting the production of 1 million cars per year in 2020, and no doubt it could be expanded well beyond that. It stands to reason that each additional Gigafactory will also be able to produce at least 1 million cars.
An accelerated production ramp to millions of cars per year could disarm hesitating incumbents of their primary advantage. It could ensure that Tesla doesn't fall significantly behind in driving data so that a competitor can't develop self-driving software with significantly better performance and safety. It could also help Tesla gain market share so that it doesn't lose to incumbents just by virtue of being unable to meet demand. The slower the incumbents move at any phase, the faster they have to move later to catch up before their advantage in production capacity is lost.
Based on their track record, I wouldn't bet on the incumbents to move fast. Based on its track record, I'm betting Tesla will.
Disclosure: I am/we are long TSLA.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.