There’s a huge challenge in getting machine learning up to speed with human learning, Hyundai Mobis’ David Agnew tells Louis Bedigian.

There are several doomsday theorists who are afraid of artificial intelligence (AI) and its potential to overthrow the planet but cars will have to learn to drive themselves before AI can threaten mankind. Never mind the Terminator-like scenarios – autonomous technology alone has proved to be more difficult than anyone anticipated. The many subtleties of driving, the very things that allow humans to get behind the wheel, are hard to replicate in a self-driving environment.

“People become better drivers as they get older until about the age of 50 or 60,” said David Agnew, director of advanced engineering for Hyundai Mobis North America, a separate company/Tier 1 supplier within Hyundai Motor Group. “You’re continually learning and getting better based on these scenarios you see. The trick is, how do you get that into your car algorithm?”

There are numerous approaches to making this happen, ranging from better AI and improved sensors to connectivity that allows cars to share and respond to data in real time. Agnew speculated that machine learning could be the answer but that would require billions of miles of testing. Few automakers have the ability run that many tests and, for those who do, it could amount to a longer learning curve than that of a human driver.

Said Agnew: “When you’re designing a system, you have a pretty good idea of how capable your sensors are and how responsive your actuators are and once you’ve got the code written you can rely on that code. The thing is: is the behaviour you programmed into it correct? Vision is a key part but my opinion is that the algorithm for the behaviour, the decisions you make to avoid accidents, are not as simple as automated emergency braking.”

Advanced driver assistance systems (ADAS) are an important part of motor vehicle safety but the rate of deadly car accidents continues to climb. Hyundai Mobis is zeroing on one problem area by developing new technology that could reduce drowsy driving fatalities. It and other ADAS features might help automakers inch closer to autonomy but it will take time to make the leap forward. Said Agnew: “To go to a Level 3 or 4 system, I think it’s an order of magnitude beyond any complexity being done currently.”

Many key challenges remain. While automakers were eager to demonstrate AV technology in Las Vegas during CES, they were unwilling to do the same a week later at the North American International Auto Show in Detroit, proving that Michigan’s brutal winter weather is still a deterrent. “People are looking for a silver bullet,” said Agnew. “I’m not sure if there is one. To me the hurdle is not a component, sensor or actuator. It’s getting to this accuracy and confidence of knowing if you can drive a billion miles without an accident.”

For now most tech companies are taking a gradual approach to AV development. Google, for example, once tested wheel and pedal-free vehicles that could drive no faster than 25mph. Agnew said that decreases the scope of the project, which is massive, but the amount of testing and analysis required remains incredibly high.

Simulation may not be enough

AV testing can be a long and expensive process but one proposed solution involves simulation. Agnew thinks it is absolutely necessary, though he is not sure it will be enough to achieve public deployment. Said Agnew: “We’ve been talking about autonomous for almost a decade now and Level 3 and 4 are still not on the road. It comes down to how safe is safe enough? Autonomous vehicles are always going to get compared to a human driver’s capability.”

According to the Insurance Institute for Highway Safety, there is an average of 1.16 deaths for every 100 million miles travelled in the United States. This includes all drivers, even those who are tired, inexperienced or intoxicated. “So a safe driver in the US really has the capability of going almost a billion miles before they have a fatal accident,” Agnew added. “If you’re going to put out an autonomous vehicle, society is going to expect it to have that type of capability. I think that is the big challenge. How do you verify that, how do you prove that?”

Agnew said it is impossible for anyone to physically drive the billions of miles necessary to prove AVs are safe. Simulations are a popular alternative but he is concerned about their accuracy. He explained: “You still have to answer the question: did that billions of miles of simulation really represent what the car will see in the real world? That’s what it comes down to. It’s not just running a bunch of simulations, it’s how you set up your simulation method and how you build a case that this is really going to drive safer than a human.”


Connected & Autonomous Vehicles

14 May 2018 - 17 May 2018, Santa Clara, USA

From vehicle electrification and infrastructure to the evolution of ADAS and vehicle automation to enhanced connectivity and new mobility models, no rock will be left unturned.