The effort to build cars that can drive themselves is reshaping the automotive industry and its supply chain, impacting everything from who defines safety to how to ensure quality and reliability.
Automakers, which hardly knew the names of their silicon suppliers a couple of years ago, are now banding together in small groups to share the costs and solve technical challenges that are well beyond the capabilities of even the largest companies. This is evident in the recently announced joint ventures by Daimler and BMW, and Ford and Volkswagen. But the changes run much deeper than just the large automakers. The also include Tier 1 and Tier 2 suppliers, which are working much more closely closely with the entire semiconductor ecosystem.
“Automakers are finding out this is too big a job for any one company, so they’re finding ways to pay for development and propagate this technology into the market gradually to help pay for the rest,” said Roger Lanctot, director of the automotive connected mobility practice at Strategy Analytics. “None of them can solve all the problems on their own.”
Some of these partnerships are in place simply to reduce cost or as a response to a short-term shortage of something very specific, according to Stan Caldwell, executive director of the Transportation Center for Safety, Traffic21 and Mobility21—U.S. Dept. of Transportation-affiliated research institutes focused on transportation safety.
Today, nearly all the major automotive OEMs have close alliances with at least one other car company, often to share the cost of the AV platform or the effort of developing an artificial intelligence (AI) pilot for the new, connected cars. The mesh of automotive partnerships helps spread costs over a wider area.
“Autonomous driving is disruptive,” Caldwell said. “Four big disruptions are actually going on at the same time—automated vehicles, connected vehicles, electric vehicles and shared-vehicle technology. The shared-vehicle is coming on first, because of Uber and Lyft, and because it doesn’t take a lot of [technical upgrades or] changes in the vehicle. Once you have a driver with a smartphone you have connectivity into the car, which makes ride hailing and sharing suddenly available. That’s why you saw all the big automakers make alliances with telecom companies to get information into the vehicle, and why they’re so big on 5G applications and vehicle-to-vehicle or vehicle-to-infrastructure communications. For automakers, everything is in flux.”
The financial situation will remain tight, particularly with slower-than expected sales of electric vehicles and slow sales of automated and autonomous vehicles. Ford CEO Jim Hackett told a Detroit Economic Club in April that autonomous vehicles are more likely to be successful on limited-access roads or walkways, at very slow speeds. “We overestimated the arrival of autonomous vehicles,” Hackett said.
Not everyone agrees with that assessment. About 40% of Teslas sold include the more expensive full self-driving technology option, according to a recent report.
Still, it’s unlikely that the auto industry as a whole will be besieged by a sudden wave of interest from consumers, who have been ambivalent about self-driving cars. A recent survey by AAA found that 71% of Americans said they are afraid to ride in fully self-driving vehicles, compared with 63% at the same time last year. The more unfamiliar a consumer was with self-driving technology, the more uncomfortable they were likely to be, although 53% said they would be comfortable in a self-driving car if there were boundaries to keep things safe—special lanes for AVs and speed limits of 25 MPH, for example. AAA survey respondents also said they were more comfortable with advanced driver-assist systems (ADAS) and were more interested in vehicles with ADAS functions—partly due to the assumption that they would be safer than humans driving themselves.
This is evident across the supply chain, as well, where some carmakers have backed off full autonomy because of slowing demand and the limits of AI technology today.
“There has been a lot of scaling back by a lot of customers,” said Geoff Tate, CEO of Flex Logix. “Earlier this year they were talking about multiple cameras and chips. Now they’re looking at one chip and one camera, so if the driver misses something the car can warn them. With multiple cameras, you can have stereoscopic vision. A single camera can’t tell you the distance.”
Tate noted that full autonomy is possible under certain circumstances, such as on a highway, but it is much more difficult in complicated neighborhoods where traffic is less predictable.
Steering wheel required?
This is reflected in the overall vehicle design, as well. Leaving the steering wheel out of a robotaxi could cause a safety crisis among AV developers.
In the U.S., a legally “safe” car is one that complies with all the requirements of the Federal Motor Vehicle Safety Standards (FMVSS)—the comprehensive set of functional safety regulations maintained by the National Highway Traffic Safety Administration (NHTSA) that covers every traditional car allowed on the road in the U.S., but it says nothing about driving automation systems or almost any other advanced electronic device, which leaves ADAS and AV technology effectively unregulated at the federal level.
To date, NHTSA has published three sets of voluntary guidelines and also promised in October that it would consider enforcing rules differently for AVs than for traditional cars, possibly to the extent of redefining rules about a “driver” or “operator” to avoid requiring that a human fill either role. It also may allow self-driving vehicles without such human-centric controls as a steering wheel or brake pedal.
Since FMVSS says nothing about AVs, however, putting a car on the road with no steering wheel would make an autonomous vehicle illegal, possibly for the first time since testing began.
The petition GM filed in January 2018 asked NHTSA for 16 exemptions to FMVSS that would vehicles with no steering wheel, brake pedal or other control surfaces to be considered street legal for testing or use as robotaxis. NHTSA had still not responded directly to the petition by July, when GM announced it would delay deployment of the robotaxi service GM Cruise officials had insisted would go live during 2019.
“The problem is that there’s no business model for safety, but sadly there is a big business based on the limitations of our current safety system,” Lanctot said. “Repairing and replacing cars is big business. Insurance is big business. But there is no business model based on making sure autonomous vehicles are safe. There is some legislative effort to address that, but the picture is so strung out and complicated it’s hard to be confident legislators will be able to solve it that way.”
In the meantime, it may fall on the automotive industry to address the safety issues.
“Intel believes that safety is the fundamental basis upon which all policy should be built,” according to from Marjorie J. Dickman, associate general counsel and global director of Automated Driving & IoT Policy at Intel. “We continue to encourage the broad AV industry to collaboratively develop a technology-neutral and transparent performance-based model for AD safety decision making, in conjunction with leading standards bodies. We urge all parties to reach agreement sooner than later, as any further delay could adversely impact U.S. innovation and deployment opportunities and unnecessarily cede ground to global markets and competitors.”
Updates to existing standards and completely new ones from well-respected groups like ISO, the Society of Automotive Engineers (SAE) and Underwriters Laboratories do reach across the gap between standards designed for functional safety of mechanical vehicles and those that address highly automated and autonomous systems in newer vehicles.
“The biggest piece missing has been this need for a formal definition of what does it mean for an automated vehicle to drive safely, because without that you have nothing to measure or test against,” said Jack Weast, senior principal engineer at Intel and vice president of autonomous vehicle standards at Mobileye.
Testing has to be broader, has to include more players and has to create both a reasonable functional base and the opportunity to be built upon, Weast said. “The biggest piece missing is a formal definition of what it means for an automated vehicle to drive safely because, frankly, without that you have nothing to measure or test against.”
He’s not alone. Test is a growing concern across the entire automotive supply chain as carmakers demand more reliability. But it also requires a different approach to test, because some of the old approaches are inadequate for automotive applications.
“The question I ask people is what is their test coverage plan,” said Jamie Smith, director of global automotive strategy at National Instruments. ” How do they know they’ve tested enough? There is probably a certain amount you’ll always have to do to figure out things, like when you approach a fire truck that might be turned a little sideways, and you can see it but the radar energy just slides along the sides of the truck rather than coming back. So the radar sees this incredibly small object in the road, not a fire truck. If you model the traces of the radar, you can see what happened, but you have to make sure to test cases like that even if they’re a part of your testing initially. A company that believes it can go from writing code and deploying it to a vehicle that’s going to be fully autonomous without doing regression testing and simulation and some level of hardware in the loop testing is not behaving in a responsible manner.”
That sentiment is echoed across the industry. “In automotive, the biggest issue is reliability,” said John Hoffman, computer vision engineering manager at CyberOptics. “You want to make sure there are zero escapes. If one customer has manufactured 25 million circuit boards, 5 defects is a huge problem.”
Until recently, the V-system testing of ISO 26262 has been the primary functional safety method for verification and validation. It will continue to play that role, according to Kurt Shuler, vice president of marketing at Arteris IP, but it will be supplemented by other types of testing such as SOTIF (safety of the intended functionality, ISO 21448).
“SOTIF was a little controversial,” Shuler said. “It almost didn’t get passed because of what I call the philosophical element. It is designed to find faults when things are working correctly, but it also finds errors that you don’t know about. The way it goes about that is a little different from the usual approach, but there are also standards coming from SAE and others from ISO, so there will be plenty of competition for this kind of challenge to be able to verify probabilistic systems.”
ISO 26262 and ISO 21448, despite the addition of test functions to let them get at some software functions remain, fundamentally, hardware verification processes, Shuler said. But even rising to identify the unknown-unknown errors, as SOTIF is designed to do, there is a limit beyond which verification procedures aren’t as valid. (See chart below for explanation of these standards.)
When you’re dealing with a system, whether it’s from the standpoint of a sensor or the standpoint of an AI/machine-learning application, it’s non-deterministic, so everything is on a distribution curve—it’s all probability because it will pass a test 999 times but not 1,000 and it’s hard to get diagnostic or forensic information to understand what’s going on inside,” Shuler said. “That’s where some of the other standards came from—SOTIF and UL 4600, the details of which are still coming out.”
It’s a difficult problem with no precedent in electronics, and there are so many possible causes of something going wrong that it becomes a challenge to identify every possible corner case.
“Every verification team is always worried about how you capture all of those corners,” said Marc Greenberg, group director for product marketing at Cadence. “Another trend that we’re seeing is a move toward more hardware-based verification. So with new chips, we’re seeing a lot more emulator solutions and simulation and a lot of cycles of large parts of a chip in conjunction with software. What’s different is that you need to understand how everything operates as a whole.”
This is particularly important in the automotive world because even if individual chips are functionally correct, they are subject to different interpretations at a system level.
“You are dealing with sensors that might see things wrongly—like a drawing on the road of a girl chasing a ball that causes the car to stop abruptly and maybe hurt passengers because it is seeing the girl in 3D rather than 2D,” Shuler said. “You’re not asking if you meet requirements. You end up asking if your requirements are right. That’s a level of ambiguity engineers aren’t comfortable with because we like to have a quantitative answer. When you are dealing with systematic errors, you’re dealing with cosmic rays switching bits and those kinds of random errors that we understand. We have a process for that, though, and there is a quantitative answer to how to find and deal with those things.”
It’s different with interactions between different systems.
“When you get up to the system level, especially with AI or machine-learning software, you’re dealing with things that are non-deterministic, so everything is on a probability curve somewhere,” Shuler said. “The questions then aren’t about whether you meet requirements, but whether you got your requirements correct. That will often get into the non-deterministic space where the answer is ‘probably,’ not quantitative.”
Fig. 1: Automotive standards changes. Source: Standards organizations/Semiconductor Engineering
Challenges To Building Level 5 Automotive Chips
The challenges to build a single chip to handle future autonomous functions of a vehicle span many areas across the design process.
How Many Test Miles Make A Vehicle Safe?
Simulation and test can improve safety, but that requires a standard framework and definitions.
How To Automate Functional Safety
Experts at the Table: With both established companies and new players clamoring to play a role in the automotive space, how is the industry moving towards automation? (Part One)
Building An Efficient Inferencing Engine In A Car
How to model a chip quickly, including corner cases.
Automotive Knowledge Center
Top stories, blogs, videos, white papers