skip to content
Article

Hyundai and Kia’s AI-based, Driver-tailored ADAS Technology

2019-12-11

Hyundai/Kia Motors have developed the world’s first AI-based partial autonomous driving technology of its kind. Named SCC-ML, the technology allows partial autonomous driving tailored to the driver’s on-road tendencies.

Cruise control is now a commonplace driver assitance feature. The early cruise control systems allowed the driver to set the vehicle speed and maintain it without action. Unless the driver were to manually adjust the speed, however, the systems had no capability to increase or decrease speed by itself.

But cruise control has since developed from a mere convenience feature to one that contributes to driver safety as well: the cutting-edge ADAS (Advanced Driver Assistance System) known as SCC, or Smart Cruise Control. At the core of its capability is the ability to not only set speeds but also to adjust the distance between vehicles by itself. At the moment, the technology is used as a core feature in the Level-2 partial autonomous driving technology.

SCC is yet again evolving now, though, by incorporating the capacity of artificial intelligence. SCC-ML (Smart Cruise Control-Machine Learning) is the product of such evolution. SCC-ML uses AI technology to grasp the on-road tendencies of the driver and offers driver-tailored smart cruise control. The cutting-edge technology will be applied to select Hyundai and Kia models. We interviewed the creators of the SCC-ML at Hyundai and Kia Autonomous Driving Commercial Development Team.

Countless Countries, Diverse Driving Conditions

The researchers who developed and commercialized SCC-ML, Manager Kim Si-Joon (left) and Senior Researcher Seo Hae-Jin (right) of the Autonomous Driving Commercial Development Team

Q. Could we start with a brief explanation on SCC-ML?

Senior Researcher Seo Hae-Jin (hereafter Seo): The existing SCC simply used the default setting for distance between vehicles or rate of acceleration, or had the driver manually set the desired levels. Even when the driver had that manual ability, though, there were only a few levels that the driver could use. For example, there were only four possible settings for distance between vehicles. As a result, it was difficult to accommodate various drivers’ preferred driving tendencies, so many drivers found SCC-assisted driving awkward. At worst, the discomfort was so great that some drivers simply disabled the system altogether. SCC-ML is the solution to this problem: by using machine learning to approximate the drivers’ preferred on-road tendencies, we provide the SCC experience that does not deviate too much from that of regular driving.

Q. In what ways did SCC-ML take advantage of AI?

Seo: Well, AI is an umbrella term for many different technologies. AlphaGo, famous for its go contests with Lee Se-Dol, is based on Deep Learning technology. SCC-ML utilizes another AI technology called machine learning algorithm to grasp the driver’s on-road patterns. When the neural network of the algorithm is complex and multi-layered, we call it “Deep Learning,” but machine learning is single-layered for SCC-ML. We considered applying Deep Learning to it, but increased complexity of the neural network can result in unpredictable errors that could become safety hazards. We figured that single-layer neural network was enough to assemble data that could realize important aspects of driving tendencies, so this simpler machine learning algorithm was chosen for the system.

Q. How do the SCC-ML’s functions work in specific?

Manager Kim Si-Joon (hereafter Kim): Sensors such as forward-facing camera and radar first gather the data from various on-road situations and then send the data to ADAS’s central control computer, which essentially functions as the system’s brain. The control computer then extracts from the input data the information needed to grasp the driver’s on-road tendencies. And then this information is stored, so that when SCC is activated, autonomous driving that most resembles the driver’s tendencies are actualized.

Q. Could we know a bit more about the control computer?

Kim: It’s the brain of the system. Again, it gets the input data from the sensors, but even data that are simultaneously gathered come in different forms from different sensors. The processing required to consolidate the different data into one useful database is actually quite difficult. ‘Sensor Fusion’ technology is applied here to resolve this issue. Using it, the computer can organize and digest the different sensors’ information into a single interface. SCC-ML uses the fused information to decide on a driving pattern and execute it on-road.

Sensor fuision functions to distill and organize the information received from the control computer.

Q. How many driving patterns can SCC-ML grasp?

Seo: Over 10,000. This includes permutations of all variables in play, which at the largest level can be divided into three categories: distance between vehicles, acceleration (how fast the vehicle accelerates), and reaction time (how fast the driver reacts to surroundings).
You might ask how it is possible to get over 10,000 combinations from just these three categories, but the variables diversify at lower levels. In the early stages of R&D, we gathered through a research contractor the driving data of hundreds of drivers, and their analysis showed that driving tendencies changed at different speed levels. For example, a driver who would accelerate rapidly at lower speeds would not necessarily demonstrate the same tendencies at higher speeds. In the same vein, there was a driver who maintained short distances between vehicles at low-speed city driving but very large distances at high-speed. These are just a few examples of variables that went into the design of SCC-ML.

Kim: We’re developing the system further to account for even more varied on-road situations. For instance, on a three-lane one-way road, if the driver is on the second lane and there are large trucks on the first and the third lane, a decision may be made to not pass the trucks. Or we could make it so that the trucks are passed at low speeds but not at high speeds. There are unimaginably many situations that could be patterned into the system. All such R&D efforts are under the objective of realizing human-like autonomous driving.

Q. How does the system “learn” the driving patterns?

Kim: The time needed to initially grasp the patterns varies, but the average is around an hour. After that, the system does routine updates to take advantage of additional accumulated data. Past data loses relative importance then, but it’s not entirely gone from the database. Essentially, as the sensors churn out the more recent data, which presumably reflects the driver’s tendencies best, the database is adjusted to give such data the priority.

SCC-ML takes into account the various on-road situations and creates over 10,000 driving patterns, thus offering a tailor-made SCC experience to any driver.

Q. Can the AI produce a new pattern that’s not included in the existing database?

Kim: We thought about this a lot in concept development. 10,000 patterns ought to be able to match almost any driver’s tendencies, but perhaps it’s not comprehensive enough. So initially we did work on the capacity to add new patterns that are identified as meaningfully different from the existing ones.
But ultimately, SCC-ML is not just a convenience feature but a safety one as well. If a new, untested pattern is applied to the system, there was a possibility that the pattern may create a safety hazard. So we discarded this capacity. If a driver’s on-road tendency is identified as very different from safe driving practices, the system matches it to the closest pattern that is also safe. Irregular driving tendencies could be a safety hazard after all.
We are considering to add more safe patterns and update it onto the system through cloudshare. Not now, though, since there are roadblocks like OTA(Over The Air; wireless update technology), but as the underlying technology matures, we predict that we will be able to update SCC-ML online.

Q. Wouldn’t SCC-ML cause discomfort in its users like SCC did?

Seo: Quite the opposite. If SCC-ML is applied, the driver won’t need to manually adjust the SCC settings. SCC-ML’s whole purpose is to automatically approximate the regular tendencies of the driver and replay it on-road. We are certain that the driver would not feel any discomfort in having the machine do for him essentially what he does without it. Of course, manual setting adjustment is still available in case it is needed.

SCC-ML is astonishingly precise in approximating the driver’s on-road tendencies.

Q. During the various testing process, how were the participant drivers’ feedback?

Kim: They were very satisfied. But as satisfaction about a function is ultimately subjective, we had difficulty turning survey feedback into more objective data.
So we approached from the data perspective. We looked not only at the subjective feedback but also at the data gathered under their real driving situations and graphically compared them to the data measured by the SCC-ML. In other words, we compared the participants’ unassisted driving data to autonomous driving data. The results, as you can see, were very positive. SCC-ML could very closely approximate the driver’s on-road tendencies.

Q. How long did the development take?

Seo: The time it took to develop SCC-ML was roughly two and a half years. That’s really short for industry standards. Taking less than three years to not only develop but apply the technology to mass-market production could be viewed as an audacious attempt.
But we have been studying AI-based tech well before we started working on SCC-ML. It just took us a bit of time to apply AI to ADAS. Our trust in our existing capacities in AI development research allowed us to trust our ability in applying it to SCC-ML.

SCC-ML technology could see the light of day so quickly thanks to the strong foundation in AI development research

Q. In the era of autonomous driving, in what ways do you see SCC-ML evolving?

Seo: Well, when the era of complete autonomous driving comes, SCC-ML might become obsolete. If humans do not drive, there won’t be data to approximate and learn from. But humans will still drive depending on the conditions even in Level-4 autonomous driving. So learning from the driver’s on-road tendencies will be possible then. In addition, the big data about driver tendencies may be useful for any number of purposes in the advent of the complete autonomous driving.

Kim: if autonomous driving were to take hold, it would need to earn the complete trust of the passenger. If the car drove differently from the driver’s tendencies, the resulting discomfort may lead to loss of trust. SCC-ML could allow drivers to grow more intimate with the idea of autonomous driving by offering a driving pattern that resembles theirs. In that sense, it could have value as a bridge technology to complete autonomous driving.
SCC-ML combines with HDA II, the technology that includes autonomous lane switching, to go beyond Level 2 and realize Level 2.5 autonomous driving. Hyundai and Kia will be applying these functions to their select new models. SCC-ML technology represents a meaningful step forward for Hyundai and Kia’s industry-leading efforts in the AI field, the crucial component of autonomous driving technology.