It may be a long wait before fully autonomous vehicles hit the road. Even semi-autonomous vehicles aren’t doing so well.

The American Automobile Association drove 4,000 miles in cars equipped with active driver assistance, averaging problems every 8 miles. AAA cited a host of problems, including driving too close to other cars or guardrails, aggressive braking, and automated steering that would disengage suddenly without sufficient notice. Cars tested were made by Cadillac, Ford, Kia and Subaru.

This is a tough problem for electromechanical systems. But for the most part, they are getting more much more reliable, even if they don’t compare to a good driver on a good day. Modeling, verifying, and testing of semiconductors has progressed with every new chip in a car. The same cannot be said for the behavior of complex AI systems — a combination of hardware and trained software — that have to navigate through unpredictable traffic, changing weather, and occasional tumbleweeds blowing across the road.

This is the impass the automotive industry needs to solve, and it’s proving much more complicated than many carmakers expected four years ago, when the race to autonomy began.
Back in 2016, carmakers were predicting that fully autonomous cars would be available by 2018, and nearly ubiquitous within this decade. Four years later, the date appears to be much further off in the distance.

There are several reasons for this. First, it’s nearly impossible to troubleshoot algorithms as they are written today. Making the hardware work properly is one thing. Semiconductor test methodologies are well understood, although improvements are still lagging due to the cost testing and inspecting 100% of the every chip used in a mission-critical application. But for algorithms, there is no simple solution. They remain opaque, and when something isn’t exactly right, systems typically overcompensate. This appears to be what is happening with AAA’s road test of ADAS systems.

The trouble is one of precision tuning and measuring what is essentially black-box technology. There is no simple way to do that. While relatively safe in comparison to human drivers, these systems aren’t as graceful as an alert human in full control of a vehicle. In fact, fine motor skills may be a misnomer when it comes to machinery.

Second, problems already are being identified in new vehicles. There will be many more problems as these vehicles are driven for extended periods of time. Testing needs to be done continuously throughout the lifetime of a vehicle, including when it is not in use. As devices age, the impact of that aging on other systems needs to be analyzed.

This isn’t a simple algorithm. It’s a full-vehicle analysis, with potential corner cases no one has ever encountered. This shouldn’t be surprising, considering a logic chip developed at the most advanced nodes has never been used outside of a highly controlled environment, and probably for good reason. But it requires much more comprehensive and ongoing analysis, which in turn requires some sophisticated technology that today doesn’t exist outside of a lab or a fab.

Third, no one has yet accounted for the impact of extreme ambient conditions on system interactions. Heat has strange effects on electronics, and no 7nm chip has ever been used for extended periods of time inside a car baking in direct sun at 120°F. At 5nm, thermal effects may be even more pronounced. Heat is known to accelerate aging, and in some parts of the world this will have a big impact on such factors as time-dependent dielectric breakdown (TDDB) and various types of memory. DRAM is known to behave badly in heat. And RF signals within and between vehicles may begin to drift beyond the filters as temperatures fluctuate and as the filters themselves begin to age.

All of these factors need to be measured, tested for and compensated for in real time. But a computer capable of handling all of these different factors would probably be well outside of the price range of most car buyers. So while the automotive electronics industry certainly is getting much better at assisted driving, it may be a lot longer until people are comfortable inside a fully autonomous vehicle where the steering wheel is optional.

AAA Full Research Report
Key findings from AAA’s closed-course testing & naturalistic driving on active driving assistance (ADA) systems
Automotive Knowledge Center
Pivoting Toward Safety-Critical Verification In Cars
Experts at the Table: Changing the automotive mindset; verification after manufacturing; security updates.
Auto Chip Reliability Opens Door To Other Industries
Auto chips are finding traction in aerospace, industrial, and even consumer applications.
Why Safety-Critical Verification Is So Difficult
Proprietary hardware makes software development more difficult; how to deal with over-the-air updates.

Ed Sperling

Ed Sperling

  (all posts)

Ed Sperling is the editor in chief of Semiconductor Engineering.