Skip to content

How to solve a problem like the semi-autonomous vehicle?

Great Wall Motors’ Dr. Vladimir Djapic speaks to Freddie Holmes about the challenge of Level 3 autonomous driving and the need for accurate maps

Many believe that driverless vehicles will be commercialised by gradually introducing higher levels of driver assistance. It is an opinion that has been nurtured by the Society of Automotive Engineers’ Levels of Driving Automation, a scale which describes six levels ranging from zero (no automation) to five (no need for a steering wheel).

Sitting in the middle is Level 3, offering a mix of both human and computer-based control of the vehicle. However, there is a lack of clarity as to exactly how and when the driver could be required to take the wheel. Ask five industry stakeholders for their interpretation, and you will likely receive six different answers.

With a Level 3 system engaged, the driver is no longer required to remain alert, despite the fact that the vehicle can only operate under ‘limited’ conditions. Somewhat confusingly, the driver must also be ready to take over if required—there is a clear conflict of interests. The argument is that the vehicle would be able to detect when a hazard is too challenging, and control can simply be passed over to the driver. This is based on the assumption that he or she is ready to take over, despite having no obligation to monitor the road. It is a recipe for disaster.

As the industry continues to pursue shared control of the vehicle, any effort to make that handover scenario safer should be welcomed

The risks became abundantly clear during the tragic event that transpired in Tempe, Arizona last year. The safety driver, by all accounts a trained operative in charge of monitoring the Uber Advanced Technologies Group test vehicle, was engaged in other activities when the self-driving system failed to register a pedestrian crossing the road.

The challenge of shared automation

Despite the inherent challenges, many automakers continue to pursue Level 3 technology. Many of the world’s major manufacturers, including Volkswagen, BMW, Volvo and PSA, are members of the L3Pilot project, a €63m (US$70.34m) effort funded by the European Commission to validate and commercialise Level 3 autonomous driving functions. Audi’s self-proclaimed Level 3 ‘traffic jam pilot’ has also been available since 2017 in the A8. With the system engaged, the driver is not required to “continuously monitor the car and can focus on another activity,” Audi states. And yet, he or she “must remain alert […] capable of taking over the task of driving when the system prompts them to do so.”

This so-called ‘handover’ scenario has left some scratching their heads. “The most difficult design challenge is Level 3 automation, which includes highly advanced self-driving capabilities that ‘sometimes’ require an occupant to take over if the automation fails,” said Chris Rockwell, Chief Executive of human experience consultancy Lextant. “It’s one of the reasons many manufacturers are skipping Level 3 and going directly to Level 4 automation.”

Michael Hafner, Head of Automated Driving and Active Safety at Mercedes-Benz Cars, warned that drivers may assume a Level 3 system is more capable than it is. “Overreliance, or ‘blind trust’, in a system can be very dangerous,” he said.

The map as a sensor

As the industry continues to pursue shared control of the vehicle, any effort to make that handover scenario safer should be welcomed. Advances in human-machine interface (HMI) aside, high definition (HD) maps have been put forward as a partial solution.

With a Level 3 system engaged, the driver is no longer required to remain alert, despite the fact that the vehicle can only operate under ‘limited’ conditions

In short, HD maps can help to prepare an autonomous vehicle (AV) as it enters the unknown. If a long-range sensor cannot accurately judge how sharp an upcoming turn may be, for example, the map can ensure the vehicle slows to an appropriate pace. At a complex intersection, that same map can ensure the front-facing camera knows which traffic lights to watch out for. As such, numerous players have dedicated mapping vehicles running on public roads around the world today.

Based in Farmington Hills, Michigan, American Haval Motor Technology (AHMT) currently has two cars roaming the Detroit area in autonomous mode, collecting valuable map data and honing the accuracy of the overall system. A subsidiary of Great Wall Motors, AHMT is leading the Chinese automaker’s push for autonomous driving in the US.

So far, tests have shown that the roads of Michigan are easier to map out than those in its home market of China, where crowded lanes can block the view of image sensors.

In June 2018, Haval—Great Wall’s SUV brand—won an award for its prowess on an autonomous cross-country course in China, but the automaker is now developing Level 4 autonomous vehicles that will operate on pre-defined urban routes.

As Mapping & Localisation Manager and Tech Lead, Autonomous Driving Systems at AHMT, Dr. Vladimir Djapic oversees how HD maps are integrated within these platforms. In his view, Level 3 systems should be approached with caution. “Level 3 is a very difficult situation,” he said. At the very least, the technology should be aware of any upcoming hazards—a road collision or a snow blizzard, for example—and pull over, rather than immersing the driver at short notice. “One of the important sensors that will enable this is the HD map,” he said, “because it provides information on different elements of the road, which can be combined with vision sensor data.”

Calming the nerves

For anyone that has used a highway pilot system, it can be a nervy experience at times. While straight roads with clear lane markings are generally handled well, today’s systems can be easily spooked and shove control back to the driver with little warning. Oftentimes, it is because a turn—whilst gradual—is sharper than expected, and the system bails out. If it knew the gradient of the corner in advance, this would help to provide context on the road ahead and not only what is in the vehicle’s immediate line of sight.

It is very difficult to detect whether a specific traffic light is red or green whilst travelling quickly; in a city scenario there may be red lights everywhere. An accurate map can provide context and tell the car where to look

“The map should definitely be considered as another sensor, but a sensor that can provide information on what is coming up,” said Djapic. The same line of thinking applies to more challenging manoeuvres, such as negotiating traffic lights without driver input. “If you don’t have a HD map, you almost cannot do the job,” he warned. “It is very difficult to detect whether a specific traffic light is red or green whilst travelling quickly; in a city scenario there may be red lights everywhere. An accurate map can provide context and tell the car where to look—or more specifically, where in the image the region of interest is.”

In fact, he suggests that the industry should strive to employ the level of accuracy required by a Level 5 system across the entire spectrum of AVs. “Otherwise, will these features really be trusted?” he continued. “Let’s not simply develop something that ‘does the job’ because it is labelled as a Level 2 technology. The better the system is, the more it will be trusted by drivers.”

The issue of trust

The progression from driver assistance to full automation should be carefully considered. Having the driver as a ‘fall-back’ in the event of a failure has already proven risky in the wild.

While Great Wall continues to investigate all levels of automation today, Djapic has his reservations around shared control of the vehicle. “A person could be behind the wheel blindfolded, yet upon a machine beeping, is supposed to take over,” he stressed. “To stress how difficult that is, who would we trust in that situation: the person that is blindfolded and has no sense of what is going on, or the machine? In my opinion, a Level 3 system should have the same level of accuracy as a Level 4 or 5 system.”

The map should definitely be considered as another sensor, but a sensor that can provide information on what is coming up

It is tough to argue the merits of a Level 3 system from this perspective, but as the industry pushes on with autonomous drive technologies, it should not ignore the support of a HD map. Alas, some believe AV developers should do just that. HD maps have come under fire by a small collective within the industry, with the suggestion that vehicles could become reliant on the map rather than the sensory input of a camera. Precise maps are “a really bad idea,” noted Elon Musk during Tesla’s investor-centric Autonomy Day in April. The same camp has suggested LiDAR will also become surplus to requirements in future.

This article appeared in the Q3 2019 issue of M:bility | Magazine. Follow this link to download the full issue.

Welcome back , to continue browsing the site, please click here