Mobileye Sets Safety-Based Rules of the Road for Self-Driving Tech
It’s been a busy year for Jerusalem-based Mobileye since Intel (NASDAQ: INTC) completed its $15.3-billion acquisition of the Israeli specialist in autonomous navigation technology.
In a recent update, Mobileye co-founder and CTO Amnon Shashua said the company has almost doubled its workforce since the buyout. It plans to begin testing its self-driving technology in California next month. And Mobileye began pushing last fall to establish an industry-wide safety model for self-driving vehicles.
Defining just what constitutes “safe driving” for self-driving cars was the primary focus of a presentation Shashua delivered earlier this month at the Intel Capital Global Summit in Palm Desert, CA. “Without standardizing the safety model, I don’t see any autonomous cars going into production,” Shashua told reporters in a subsequent press briefing.
Mobileye was founded in 1999 to commercialize computer vision technology that Shashua developed as a professor of computer science at Hebrew University. Intel became a strategic partner, and when the chipmaker announced its buyout in March 2017, Mobileye was the auto industry’s leading supplier of advanced driver assistance systems (including collision avoidance technologies).
Shashua said Mobileye has partnerships with 27 major automakers, primarily in the United States and Europe (with another 20 automakers in China). It has been road-testing its driver assistance systems since 2007, and has collected over 200 million miles of driving data. He argued that the scale of the data already collected gives Mobileye a big advantage over startups that claim to have developed autonomous navigation technologies capable of operating all safety-critical driving functions while monitoring the surrounding environment (i.e. Level 4 of the five levels of autonomous driving).
Shashua voiced skepticism of such claims in his media briefing, saying startups have not amassed enough data to definitively evaluate their autonomous technologies in the range of conditions that Mobileye has.
“We’re talking about something that is safety-critical. There is a big gap between something that kind of works and really works,” Shashua said. “We are in a field where good enough is not good enough. In many domains, you can buy a product which is good enough. If it doesn’t work, nothing bad happens. So you reboot your smartphone. But since we’re talking about safety-critical, there is no such thing as good enough.”
With about 35 fully autonomous cars (i.e. Level 5) being built with Mobileye technology, Shashua said, Mobileye plans to begin test-driving in California next month.
Since Intel’s buyout, Shashua said Mobileye’s workforce has mushroomed from 800 people to roughly 1,500. “Mobileye is an independent company, but over time, more and more Intel people are joining,” he explained. “One of the things we are getting from this acquisition is very fast scaling up in terms of manpower.”
Such growth has been “very difficult,” Shashua added, “because the kind of engineers we hire are very specific engineers”—the company specializes in computer vision and machine learning. Beyond the engineering challenges, though, Shashua said Mobileye has been working on all sorts of elements that go beyond developing technology—especially in terms of a formal safety model that Mobileye calls Responsibility-Sensitive Safety, or RSS.
Given that self-driving cars will be sharing the roads with human-driven vehicles for decades, Shashua said the underlying premise of RSS is that accidents involving self-driving cars are inevitable. He argued that such a model is needed to quickly determine whether or not the autonomous vehicle was at fault, “because when there is an accident, who is liable? It is not the passengers. It is the suppliers of the technology and the [auto] manufacturers…You need to have a model that clarifies the notion of blame, and formalizes the notion of what is dangerous, formalizes the notion of who is responsible.”
So, Mobileye has rolled out its formal safety model as the first step in establishing rules of the road for autonomous vehicles. Shashua said the model uses mathematical formulas to formalize a “common sense” approach to safety. The rules establish concepts of priority in various driving scenarios, setting parameters for maintaining speed and distance from other vehicles. For example, Shashua used a video demonstration to show how such rules would guide a self-driving car as it merges with freeway traffic.
“Merging into dense traffic requires significant ‘negotiations’ among [human] drivers,” he said. A self-driving car must emulate the same sort of assertiveness. If it waits for an opening, it can create a bigger problem by blocking other cars also attempting to merge.
With RSS, Shashua said, “We define in advance, what it means to be in a dangerous situation. What does it mean to have a proper response, and then guarantee that the actions our car would take would never cause an accident. So that if there is an accident, all you need to show is that you did not have a sensing mistake… If you show there is no sensing mistake, then the car’s actions cannot cause an accident. Therefore, the blame of the accident is on the other driver.”
Shashua said the company laid out its concept for RSS in an academic paper published last year, and it has been working with transportation regulators in a bid to establish RSS as an industry standard. In a speech last October in South Korea, Shashua called on industry leaders and policymakers to collaboratively construct standards that definitively assign accident fault when human-driven and self-driving vehicles inevitably collide.”
Shashua’s presentation was perceived in some quarters as an attempt to avoid liability. Rather, he told reporters the model is intended to address some basic questions about autonomous navigation, such as, “What does safety mean?” and “Can a self-driving car guarantee ‘safety?’”
“Without such a model,” Sashua said, “autonomous cars will remain a science project. It will never reach mass production.”