After a decade of bold claims and massive investment, what do stakeholders see as the biggest hurdles for automated driving – and how have these changed their approach? AAVI recently caught up with experts across the ADAS & autonomous vehicle sector to share their thoughts.
CHALLENGE 1: SCALABLE SOLUTIONS
The biggest near-term challenge is finding affordable, useful applications for automated driving. Addressing CES in January 2023, Mobileye CEO Amnon Shashua said OEMs are increasingly aware of opportunities to provide conditional automation using already-available technology, such as its SuperVision driver assist system. The company is bridging the gap between ADAS and automated driving using modular components, and incrementally expanding the operational design domain (ODD), beginning with highways.
Mercedes-Benz is taking a similar approach. Its SAE Level 4 valet parking was approved for use at Stuttgart Airport last year, and its Drive Pilot Level 3 highway assistance is certified in Germany and Nevada.
Built on the foundations of ADAS hardware, Drive Pilot adds centimeter-accurate HD maps, lidar, a rear-facing camera and microphones to detect emergency service vehicles. It can maintain speed, distance and position within a lane and perform evasive maneuvers at up to 60km/h, and Mercedes claims 90km/h is possible using the same hardware. The company’s longer-term target of 130km/h relies on higher-definition lidar.
Dr Egil Juliussen, senior partner and principal analyst at VSI Labs, says innovation is driving down the cost of systems: “Lidar is moving toward a few chips, which will solve price issues. By 2030 it is likely that all AV sensors will cost around US$500 including full camera, radar, lidar and thermal cameras. Lidar is moving to frequency modulated continuous wave (FMCW), especially for long-range, and radar is moving to 4D imaging radars, due to feature benefits.”
Wayve is also aiming for improved scalability. Kaity Fischer, vice president of commercial, says expensive sensors, HD maps and rules-based coding have made new vehicles, locations and applications prohibitively expensive and time-consuming to deploy. Instead, the company is using human drivers to demonstrate maneuvers that a machine learning algorithm can then apply in other locations and scenarios.
“Our AV2.0 approach removes the need for HD maps and programmed rules, so vehicles powered by Wayve’s autonomous solution are not restricted to predefined routes. Therefore, they can drive in ‘new’ areas with minimal preparation and infrastructure,” Fischer explains.
“This makes it highly attractive for commercial deployment,” she continues, “as few transportation services operate on fixed routes. Additionally, Wayve is taking a vision-first approach [using cameras], which means the cost of our hardware system is palatable for transportation customers.”
CHALLENGE 2: EFFICIENT TESTING
Global road conditions present near-infinite edge cases, and extensive testing and validation are vital. Dennis Winslow, operations manager at Intertek’s American Center for Mobility (ACM), a custom-built connected and autonomous vehicle proving ground (pictured, above) near Detroit, Michigan, says the industry is learning the potential and limitations of the available technology but that a new approach is needed.
“Data acquisition to support this process can become a bottleneck in the development process,” he says. “Specifically, more effort is required to address all environments, scenarios and identification of potential fringe case scenarios to address false positives/negatives. However, there are many solutions being utilized, including mass data acquisition performed by on-road fleets, HIL/SIL solutions, simulation and augmented reality, and advances in intelligent transportation systems (ITS).”
OEMs and Tier 1s are helping to create that data. Mobileye’s REM system is harvesting 28 million miles of data per day from forward-facing cameras, optimized to transmit at a rate of 10Kb per kilometer, which enables it to build detailed virtual environments for testing. Similarly, working with Nvidia, Mercedes-Benz is gathering data from vehicle sensors and using this to generate and test rare and hazardous scenarios in the Drive Sim platform.
The car maker’s chief software officer, Magnus Östberg, explains, “Our vehicles are able to collect up to 300PB of data every year to further train our deep neural networks. And our adaptability in vehicle AI automatically chooses a fraction of this data that is the most useful for improving our software stack. Our next-generation automated driving [technology] will learn about and handle infrastructure changes and continue to improve throughout its life.”
However, TÜV Süd’s global head of autonomous driving, Christian Gnandt, notes that this essential process also has inefficiencies, as it requires millions of kilometers of driving and can’t address every possible scenario. He believes stakeholders require objective, industry-wide criteria measuring the criticality of those events.
“These metrics should reflect the relevant pass/failure criteria metrics for each use case and adopt threshold performance levels for validation and verification. With this foundation, a criticality coverage analysis can be conducted, allowing regulators to more effectively evaluate whether a certain technology has sufficiently demonstrated safety and roadworthiness prior to certification,” he says.
CHALLENGE 3: IDENTIFYING EARLY ADOPTERS
Passenger cars won’t necessarily be one of the earliest markets for higher levels of longer-range autonomous technology. The California Department of Motor Vehicles Autonomous Vehicles Program has issued permits for testing features since 2020, and deputy director Dr Bernard Soriano says rules have been adapted to support markets that are closest to commercialization.
“California’s regulations initially did not allow for commercial vehicles under 10,000 lb [4.5 metric tons]. As the technology was being developed, it became clear that there was going to be a space where the timelines were a little more accelerated, so the need for companies to be able to test [those vehicles] became clear. The technology was being developed much more rapidly than anyone anticipated,” he explains.
Waabi is also focusing on heavier, long-haul freight. It is developing an AI-based, scalable plug-and-play system for trucks and claims state and federal regulators have realized there’s a market need for it. The company believes that automation can help alleviate the current labor shortage of 80,000 truckers, bolster supply chains and reduce the 40,000 annual fatalities on US roads.
Dustin Koehl, Waabi’s head of transportation, comments, “The urgency to deploy AV in trucking grows every day. At the same time, the path to deployment in long-haul trucking scenarios is less complex than in urban environments. This allows us to realize the benefits of our technology safely and successfully.”
Embark Trucks has similar aims, targeting 2024 deployment in a market that chief commercial officer Sam Abidi says offers a reduced technical challenge and large societal impact. The company’s approach is enabled by transfer points across the country, where freight can be moved from driverless long-haul trucks to manned first- and last-mile delivery vehicles. That infrastructure has been rolled out across the US Sun Belt, offering foundations for coast-to-coast shipping.
“Automated trucking allows developers to avoid solving some of the more complex scenarios that take place inside city centers, such as intersections or close pedestrian encounters, by confining trucks to the highways and side streets that connect transfer points,” comments Sam Abidi, CCO, Embark. “It also has the potential to meaningfully reduce existing supply chain bottlenecks by answering the chronic driver shortage with cost-effective transportation capacity that is capable of safely running 24/7.”
CHALLENGE 4: DEVELOPING INFRASTRUCTURE
Vehicles are likely to become ever-larger contributors to the Internet of Things, but a reliance on nascent and expensive connected infrastructure could stifle wider deployment. To avoid those constraints, Embark Trucks is using lidar and camera sensors to respond to real-time scenarios such as roadworks and lane changes instead of relying on potentially outdated HD maps and non-existent connected infrastructure. Recognizing critical road users is part of that process, and last year the company validated features that enable trucks to interact with emergency services.
Embark’s CCO, Sam Abidi, explains, “This required our engineering team to train Embark-powered trucks to identify emergency vehicles via lights and other cues, and then respond accordingly by pulling over safely onto highway shoulders. They also developed an interaction procedure to enable any law enforcement officer to safely stop, approach and receive information from an autonomous truck intuitively and without any additional equipment.”
Egil Juliussen of VSI Labs foresees a longer-term use for C-V2X technology but not for at least a decade – aside from smartphone-based applications for pedestrian safety. In the meantime, he says, there are other issues to address.
“The most useful infrastructure is road markings for lanes, exit and entry lanes. That is the ground truth for the vision systems used in ADAS and AV systems. Road design has an impact on how far the vision systems can see lane markers and determines driving speed and maneuvering. Good base maps are needed for augmentation,” he explains.
“We have become experts in how important lane markers are for both ADAS and AVs. In many ways, lane markers are becoming a key technology for improving driving vision systems. Some in the US Department of Transportation understand this and are getting data to tackle this problem. VSI is working with multiple states in the US to gather lane marker data and analyze and find weaknesses and solutions. We use camera, radar, lidar and thermal cameras and drive thousands of miles to get an understanding of what is important for visions systems in ADAS and AVs.”
Waabi uses cellular connections for dispatching and monitoring, but its system is also designed to operate without relying on digital infrastructure. Sam Loesche, head of policy and public affairs, says that although clear signage and markings are beneficial for all road users, Waabi Driver can cope without them.
“Even in the case of an obscured road sign or inconsistent lane markings, the Waabi Driver still has an advantage over human beings. We also utilize a treasure chest of information including high-definition maps and endless hours of experience on the very roads on which we are operating to ensure we can operate safely in any situation,” he comments.
CHALLENGE 5: SUPPORTIVE REGULATIONS
Consistency of regulations is vital for non-geofenced applications. Bernard Soriano of the California DMV believes there’s a need for harmonization at a federal level to avoid a patchwork across the USA, and adds that rules should enable regulators to keep track of how the technology is developing. In California, operators have to submit reports to the DMV, which offers a useful oversight.
“One of the things we receive here are crash reports from the companies that are testing their technologies. Oftentimes, what we see on the crash reports are situations where the autonomous vehicle is obeying the traffic laws and that obeying a traffic law does not comport with what we as humans are used to,” he says.
“We’re going to see a mixed fleet on our roadways for a long time, and that driver behavior needs to be taken into account when you talk about the wide-scale adoption of the technology.”
TÜV Süd’s Christian Gnandt also sees a need for harmonization of regulations and standards, combining virtual and physical testing, functional safety and cybersecurity. Those conversations are taking place: the organization is a founding member of the International Alliance for Mobility Testing and Standardization (IAMTS), a global association that is creating the dialog needed to share test information and differences in global regulations, which Gnandt believes will accelerate the development of autonomous vehicles.
“Legislation should take the lead – legislators should create a framework that enables technology development. In Germany, a new law for autonomous driving is being put into force. The accompanying regulation is in the final stage and TÜV Süd is excited to start the first projects with the target of getting approval for Level 4 vehicles according to AFGBV (Autonomous Vehicles Approval and Operation Ordinance) and EU Level 4 autonomous driving law,” says Gnandt.
A supportive regulatory framework for AV testing is crucial to create a favorable business environment, adds Wayve’s Kaity Fischer. Announced in 2022, the UK government’s Connected and Automated Mobility 2025 roadmap identified a £42bn (US$50.4bn) boost to the economy that AVs can create by attracting investment and creating jobs, and Wayve is seeking further clarity about how this will shape up.
“The UK risks losing this market if the government does not bring forward legislation to allow the commercial deployment of self-driving vehicles by 2025. Unfortunately, progress has stalled since the government committed to legislation in the 2022 Queen’s Speech and we are yet to see a new timeline for primary legislation,” Fischer comments. “Legislation is only the first step to commercial deployment on UK roads. We will also need to fulfill new regulations to ensure that a safety and assurance framework is in place to build public confidence and trust in this technology.”