Startup suggests verifying self-driving cars like computer chips

A self-driving car approaches an intersection. The light turns yellow. Another vehicle trails closely behind.

Should the self-driving car speed through the intersection? Hit the brakes? What if the tailgating vehicle is 30 feet behind the autonomous one instead of 20 feet? What if the roads are wet? What if there are pedestrians near the road?

Speed, distance, road conditions. The number of permutations from a single general scenario are almost endless.

"There are hundreds of millions," said Ziv Binyamini, co-founder and CEO of Foretellix, a Tel Aviv, Israel, startup that argues that driving millions of test miles on public roads is a lousy and time-consuming way to seek encounters with all those scenarios.

The company wants the auto and tech industries to reconsider the way self-driving systems are vetted and validated. Foretellix says it has pioneered software, called Foretify, that can generate tens of thousands of variants from a given driving scenario. Starting Monday, May 13, company executives are showcasing the technology in public.

Whether in simulation or physical driving, automakers can use this software to test the mettle of their self-driving systems. Then, Foretellix can cobble results together from all platforms and deliver a scorecard or grade.

Such a system might be useful for companies validating their technology or for regulators, who are only now beginning to ask broad questions about how they might better understand and evaluate the competence of self-driving systems.

"You can see how you're doing both in a physical or simulated environment," Binyamini said. "How did you do against this overall test plan? How well did you cover the yellow-light scenario? Did you get 93 percent of them, but not do well in the rain or at high speeds with a pedestrian jumping in the middle? We can give very accurate measurements."

Building more comprehensive scenario libraries in rapid fashion might be one benefit for automakers — Foretellix says it's already working with one automaker in North America and another in Europe. Figuring out where performance doesn't meet muster helps focus development efforts.

The company, founded in February 2018, draws its roots from the semiconductor industry. Binyamini previously was vice president of engineering for Verisity, a company purchased by Cadence Design for $315 million in 2005. He has developed verification tools at Intel, where he worked with fellow Foretellix co-founder Gil Amid, who worked at Intel for 30 years in a variety of verification roles.

In terms of measuring safety, Foretellix isn't the first to propose a means of setting some benchmarks or metrics beyond the well-known miles driven and corresponding number of interventions required by human safety drivers. For one example, Israeli supplier Mobileye, part of Intel, has proposed what it calls Responsibility-Sensitive Safety, a mathematical model to ensure autonomous vehicles follow certain rules to keep them within a performance envelope that ensures they never cause a crash.

Binyamini says that's a good start and that Foretellix will integrate an open-source version of Responsibility-Sensitive Safety into its software. But ultimately, he says, that's only a start and not enough.

"It's only some rules," he said. "It's not everything that needs to be checked. If I stay within this envelope, I still need to check that the car stays within the safety envelope within all conditions. One test doesn't mean it will be OK in billions of other scenarios."