The Mobileye Safety Methodology for Fully Autonomous Driving

The Mobileye Safety Methodology for Fully Autonomous Driving

Our methodology for safety architecture in autonomous driving builds on foundational principles and industry standards to mitigate different types of risks.

Prof. Shai Shalev-Shwartz and Prof. Amnon Shashua

When it comes to autonomous driving, how safe is safe enough?

To date, even with thousands of self-driving vehicles on the road worldwide, and with AV commercialization on the near horizon, this question remains unsettled. One key benchmark so far has been mean time between failures (MTBF) -- basically, can a self-driving system drive better than an average human in terms of frequency of accidents or harm? This metric is easy to grasp and measure, but under close examination its limits become clear. Human statistics are heavily impacted by illegal and careless behavior; a self-driving system can't get drunk or text someone. More importantly, humans aren't just measured by accidents, but by avoiding reckless behavior - taking a duty of care toward themselves and other road users to avoid unreasonable risks. This is why we argue that while a high MTBF is critical, it is not sufficient to demonstrate a safe SDS.

After years of meaningful progress, both Mobileye and the industry at large now have a clearer picture for safety requirements around fully autonomous driving. Today in a new paper, we unveil a framework designed to deploy safe, self-driving systems at scale building upon two key principles:

• Overall mean time between failure of the system should be at least as good as human statistics.
  
  
• The system should eliminate unreasonable risk, where the self-driving system provider is transparent about the boundary between reasonable and unreasonable risk.

The first requirement addresses the greater good baseline - adding self-driving vehicles to the road must not cause more harm than the status-quo of roads with human-driven vehicles, as measured by MTBF. However, this statistical measurement is not sufficient and therefore we complement it with a requirement of eliminating unreasonable risk - a sort of adjusted MTBF that incorporates the principles of transparency, accountability, and adherence to robust safety standards.

The difficult parts are to define the boundary between "reasonable" and "unreasonable" risk in a rigorous manner as well as to set up the methodology for eliminating unreasonable risk. For the technical details on how we do it please read our paper.

Share article

To stay up-to-date on the latest at Mobileye

Related News

Press Contacts

Contact our PR team