Predicting the future is hard. Yet many people talk about self-driving cars as if they will become reality in a few years. Unfortunately, the current reality is that we are far away from commercialization of fully self-driving cars. Google is the closest, yet its self-driving car technology has some serious limitations:
- It requires an attentive human driver to drive safely. This largely defeats the point of a self-driving car.
- It doesn’t work if there is heavy snow or rain.
- The car only works in areas with special 3-D maps, which are currently expensive to create.
- The system can’t handle construction zones.
- Because they drive in a non-human manner, the cars get rear-ended more often than human drivers.
- There are other situations where the cars may have problems – left turns without a light and heavy traffic, potholes, pulling aside for emergency vehicles, obeying directions from a police officer, ice on the road surface, cyclists doing a track stand, etc. etc.
I find it interesting that so many people have been sucked into the idea that self-driving cars will be an imminent revolution that will disrupt our lives. Mostly, there are many people who want to believe that technology will disrupt our lives in a positive way. There is no differentiation between technologies with major technical obstacles (e.g. artificial intelligence, machine learning) and technologies with few obstacles (e.g. cloud computing, social media, smartphone apps, over-the-top video, etc.).
Partly, this is because certain companies have been putting on dog and pony shows. Google, Tesla, and Uber have released videos of cars driving themselves. Google and Uber have taken journalists on rides in self-driving vehicles (Uber’s trips often involve several interventions by the human driver behind the wheel). It might appear that the technology is close to 100% functionality. But it can be hard for people to understand that these cars are driving themselves only in the easiest situations. Or, these cars are more dangerous than they appear.
There are major technical obstacles. For example, computers are not good at detecting objects in the vehicle’s path based on machine vision. Computers can do it, but the accuracy is far inferior to human beings. In general, object detection is one of the many simple tasks where computers lag behind human beings despite very smart people trying to crack the problem for many years (e.g. optical character recognition). If you’re not convinced, you can go on Youtube and find videos of Teslas smashing into stationary cars (here is a fatal crash in China and here is a non-fatal crash where the side of the car made contact with a stopped vehicle). Or listen to the co-founder and CTO of the key supplier behind Tesla’s Autopilot describing the system as unsafe to his academic peers.
Instead of relying on machine vision, Google’s self-driving cars use something called LiDAR to build a 3-D model of the vehicle’s surroundings. That 3-D model is compared with a special 3-D map of a city that is constructed beforehand by Google (at a much higher resolution than Google Maps). By comparing the two, the computer can figure out what objects are part of the road and what objects aren’t fixed to the road- cars, cyclists, pedestrians, obstructions, traffic cones, etc. This deals with the object detection problem. The problem with LiDAR is that (A) LiDAR has problems with rain and snow and (B) the system breaks down when the real world does not match the special 3-D maps needed to make the system work. If a temporary red light is setup around a construction zone, the car may potentially drive through the red light without stopping.
What some experts are saying
John Leonard, a MIT professor that was part of a team that participated in the 2007 DARPA Urban Challenge for autonomous vehicles, is quoted in a MIT Technology Review article:
Leonard is restrained in his enthusiasm for the commercial trajectory that autonomous driving has taken since then. “Some of these fundamental questions, about representing the world and being able to predict what might happen—we might still be decades behind humans with our machine technology,” he told me. “There are major, unsolved, difficult issues here. We have to be careful that we don’t overhype how well it works.”
Ryan Eustice, who also took part in the DARPA Urban Challenge, is also reserved:
Here’s the bad news: currently the most advanced systems such as Tesla’s Autopilot function that allows hands-free operation in certain situations are not as safe as one death per million miles driven, according to Olson and Eustice.
Mobileye is one of the leaders in actually commercializing a self-driving car product, which is a key component in Tesla’s Autopilot functionality. As mentioned previously, its founder and CTO Amnon Shashua has described his own company’s product as “unsafe driving”. Note that he has an incentive not to say such things because Tesla is a Mobileye customer. (To be fair, he is an optimist. His presentation does lay out a roadmap towards fully self-driving cars assuming that multiple breakthroughs in technology do occur.)
Elon Musk is reckless and is getting his customers killed
The ugly divorce between Tesla and Mobileye has to do with Tesla’s mixed messages over Autopilot. The name Autopilot gives consumers the false impression that the car is self-driving. Because the system works most of the time, consumers may be lulled into a false sense of security. However, the system has a major flaw/limitation because it will occasionally fail to detect other vehicles on the road and drive into them. The system really does require an attentive human driver. If Tesla had described the system more accurately as “cruise control on steroids”, then perhaps Tesla drivers would pay more attention. However, Elon Musk has been telling the world that Autopilot would have “saved” lives if it was more prevalent and that self-driving cars will be a reality in a few years. Mobileye was not happy with Tesla’s recklessness in communicating with drivers.
Note that Amnon Shashua made his unsafe driving remark before the first Autopilot fatality. Unfortunately, it turns out that he is correct about the dangers of Tesla’s Autopilot.
What human beings want to believe
Some investors like to invest in companies that will benefit from emerging secular trends. They want to spot the next 20-bagger. To some degree, some CEOs tell shareholders what they want to hear. They might tell the self-driving car story to boost the share price ahead of an equity raise, or simply to maintain their image so that investors leave them alone and keep the CEO’s job safe.
And partly, there is some ego and hubris involved. Much like how most drivers believe that they are above average drivers, CEOs may believe that they are somehow special and that they will be able to lead their company to self-driving greatness. CEOs are human beings too. It’s reasonable to expect them to be overconfident in their abilities and to ignore annoying facts about reality. It’s how businesses get started– the founders generally ignore the riskiness of startups and the low odds of success. (That’s how I started my own image processing business.)
And partly, some people want to believe in technology. In the same way that Tupac Shakur and Elvis Prestley fans want to believe that these musicians are alive, some human beings will believe what they want to be true. Some people want to believe that technology will disrupt our lives and make the future look very different from today. They want to believe in buzzwords like “machine learning” and “artificial intelligence”. The reality is that those are areas of computing that don’t work very well- computers lack intelligence and aren’t very good at learning. Unfortunately, the successes in those fields have been limited to narrow niches like playing Go. Of course, there are areas of computing that do work well- controlling machines (e.g. traction control on cars, robots that build cars), cloud computing (which has been around since the 60s), search engines, traffic maps, etc. However, there is little distinction between things that computers are not very good at versus things that computers have been proven to be good at. Instead of seeing computers as useful tools with some limitations, they are viewed as magical devices that will transform every aspect of our lives.
Successful applications of computing in vehicles
Currently, fully self-driving vehicles have been successfully deployed around the world. Some subways systems have self-driving trains. The obstacle towards wider adoption of self-driving trains is the need to detect objects on the track such as fallen trees, cars stopped at a railroad crossing, and unauthorized access from passengers. Currently, human beings are much better at this task than computers. The workaround for this problem is to close the track. Barriers (platform-edge doors) prevent passengers from going onto the track. Because it is expensive to retrofit older stations with the technologies that enable self-driving trains, many subway systems continue to use human operators.
Subways are a domain where it is possible to sidestep the challenges that self-driving vehicles face. It is likely that self-driving vehicles will slowly see more applications in domains where the technical challenges are lower or can be side-stepped. We may (or may not) see computers automate easier driving situations such as highway driving as well as stop-and-go traffic.
While some may see Mobileye as a pure play on the self-driving car (e.g. their IPO roadshow video has a woman putting on makeup in a moving car), Mobileye is a highly-profitable company that makes intellectual property for computed-assisted driving (or ADAS / advanced driver assistance systems). The company may continue to be highly profitable even if fully self-driving cars do not become a reality. The company is very commercially-minded as it has worked on products involving cheap hardware (video cameras, not LiDAR) and paths to near-term commercialization.
I personally see self-driving cars as a long shot. While it’s possible that multiple scientific breakthroughs are achieved, the odds are low. However, computer-assisted driving is more interesting. I’m optimistic about computers being able to make driving safer or simply less annoying. While I have no plans to invest in Mobileye, I can see their profits growing.
Tesla is a company that I’m skeptical about, though I don’t think I have any good unique insights on the company. Certainly I do not believe that Elon Mush is a technical visionary. He is putting his customers at risk with cruise control on steroids. The SolarCity scheme is silly because residential solar is not economic versus utility-scale solar without subsidies. (I’ve also lost money shorting Tesla and apparently I have not learned my lesson.) We’ll see how it plays out. It does not strike me as a great short.
*Disclosure: No positions in GOOG or MBLY. You should assume that I may be short TSLA.
Skepticism about self-driving cars
MIT Technology Review has a lot of good articles on self-driving cars. You can use Google to search site:technologyreview.com if you’d like
Self-driving vehicles in other domains