Is a complex mathematical treatise a sufficient substitute for real-world safety testing? Israeli company Mobileye, owned by Intel, is seeking to show that its unique approach to self-driving cars is more practical and effective than methods employed by companies like Waymo. Their CTO, Amon Shashua, wrote in a blog post in October: “We target a vehicle that gets from point A to point B faster, smoother, and less-expensively than a human-driven vehicle; can operate in any geography; and achieves a verifiable, transparent 1,000-times safety improvement over a human-driven vehicle without the need for billions of miles of validation testing on public roads.” Is this likely, or are these bold claims by the unorthodox company?
Mobileye’s Self-Driving Cars
While most self-driving cars rely heavily on lidar and radar systems in order to read their surroundings, Mobileye favors an all-camera set up. According to Mobileye, their self-driving cars use cameras to make a 3D model of the real world and then use software to decide what to do with that information. This is in contrast to other, more conventional vehicles in this same industry that don’t rely on cameras alone.
Sensor Fusion as An Alternative?
“Sensor fusion,” a process by which a self-driving car could be made to use both cameras and lidar, is an end goal for Mobileye. Whether their proposed method of designing these cars will work or not is uncertain. Recent demonstrations have shown Mobileye-powered cars even speeding through red lights, which is more than a little troubling.
It would seem that Mobileye has banked heavily on sensor fusion being a foolproof way to ensure their vehicles are safe. This rides on the assumption that cameras will only fail in situations that lidar will succeed in, and vice versa. However, this assumption itself seems erroneous: what about situations where neither sensor works? Not to mention the hassle of coding to get the two sensor types to cooperate. And, as every programmer knows, the more lines of code something has, the more likely it is to have bugs.
Responsibility-Sensitive Safety
Where Waymo and similar companies prefer hours of real-world road testing for safety, Mobileye has shown a desire to prove their cars are safe via math and statistical models. Mobileye has even gone so far as to create a Responsibility-Sensitive Safety, or RSS, model that attempt to mathematically quantify all the ways a vehicle could be at fault in an accident. By putting this RSS model in all their vehicles, Mobilieye claims it would then be impossible for them to be at fault in an accident.
The logical gaps in this are pretty clear: the model can’t possibly cover every scenario that could occur. Or, even if it does, it can’t be implemented by every self-driving car all the time. The real world simply doesn’t allow for clean, clinical, one-hundred-percent-accurate models. So why would Mobileye push for this shortcut instead of just safety testing their technology?
Business Model Influencing Safety Model
Waymo, as an example, is making their self-driving cars in order to operate a driverless taxi service. They aren’t beholden to any specific deadlines and they aren’t selling their technology to anyone else. Mobileye, on the other hand, is: they sell parts to many major car manufacturers. To this end, it is important to them to have marketable, safety-tested self-driving technology quickly. Spending excess time on real-world testing must seem a waste to a company on a deadline.
This doesn’t even account for climate and surroundings. Mobileye’s parts have to work in nearly any climate, whether that be snowy, sunny or rainy. Waymo, by contrast, has the luxury of rolling out their taxi service wherever they please. This allows Waymo to choose ideal conditions, where Mobileye has no such luxury. While this doesn’t excuse lax safety standards, it at least explains them.
Here’s hoping that their self-driving technology is up to snuff by the time it goes live. Mobileye wants to make money, certainly, but if it comes at the cost of more motorists or pedestrians being killed by software or hardware malfunctions, then it’s not worth it.
