The success of failure of self-driving cars — that’s whether you will eventually be able to surf the web, chat with friends, read emails, or even take a nap while your car pilots itself down the road — comes down to code. Every major automaker employs an army of programmers tasked with virtually the same goal: Write the code that keep a car rolling safely down the road to its destination.
The problem is that eventually, they’ll all share the same road. Which begs the question: If self-driving systems can’t talk to one another, can multiple systems safely navigate roadways without conflict?
“Everyone going into the first DARPA Urban Challenge in 2007 had similar concerns,” recalls Bryan Salesky, who helped build the self-driving Chevy Suburban that navigated the complex urban course to victory. Today he’s the CEO of Pittsburgh, Pennsylvania-based Argo AI, the company Ford commissioned to build the automated driving system it will roll out in 2021.
“It was the first time that multiple driving systems were all operating on the same test course and interacting with one another and needing to obey the same rules of the road and so on,” recalls the high-tech entrepreneur. “We were delighted to see the interactions were surprisingly human-like. The advanced systems were able to interact with other robotic traffic without any issues.”
Today, the “challenge” is much different. It’s not about whether we can achieve autonomy, but rather which company will produce the first vehicle and win first-mover advantage in the marketplace.
AI developers and safety advocates think a single driving system is the safest, most efficient pathways to autonomous driving.
The race has also moved from a relatively small, highly controlled test course to the real world. The geography is more diverse. There are more self-driving cars competing for the prize. And from pedestrians to bouncing balls, the obstacles are as unpredictable as ever. Not to mention, self-driving vehicles will have to share the road with those driven by your mom, dad, Aunt Ida, and every other human that chooses to pilot their vehicle.
In a nutshell, there is more risk. If a driving system failed during the Urban Challenge, the vehicle would do little damage to person or property. That cannot be said about an out-of-control autonomous vehicle (AV) in downtown Pittsburgh, for example. Search “Uber and pedestrian,” if you have any doubts.
So, can competing self-driving systems still coexist?
Unfortunately, there is little consensus on this subject. Some AI developers and safety advocates think a single driving system — a unique set of code that manages all driving responsibilities — is the safest, most efficient pathways to autonomous driving.
“If we had a single system with clear expectations, clear deliverables, clear failure modes, and outcomes there would be a greater level of transparency and understanding,” says Deborah Hersman, CEO of the National Safety Council. “There would be no more race to gain a first-mover advantage in the driverless evolution, and every developer would be building around a sure thing.”
Others think a standardized set of rules for how AVs should behave in different situations is the best approach. Relative newcomers to the AV race, NuTonomy and Voyage are in favor of such a strategy. Both have published papers that provide a framework for how a self-driving vehicle should behave in a wide range of circumstances, including pedestrians being in the road, nearby cars reversing, and arrival at a four-way stop, which they hope other developers will use as the cornerstone for their programming efforts.
Voyage has made the company’s internal safety procedures, materials and test code all open source to provide “a foundational safety resource in the industry.” CEO Oliver Cameron Cameron told Ars Technica that after the deadly Uber collision in Tempe, Arizona, he had to spend a lot of time calming people down, telling folks it was an isolated incident. “But the truth is that everyone in the industry is reinventing the technology and safety processes themselves, which is incredibly dangerous,” Cameron told the magazine. “Open source means more eyes, more diversity, and more feedback.”
However, some experts see a serious risk to the one-code-fits-all approach. “It locks everyone into a single system that might not allow for future innovation or change going forward,” warns Argo’s Salesky. “Not everyone is solving for the same reasons or solutions.”
“[Sharing data and work] is the only way to produce the safest, best-driving AI possible.”
Instead, Salesky and others believe there is utility in having multiple efforts, multiple approaches to tackling the issue: “Design diversity is one of the pillars of building robust and fault-tolerant systems that can better respond to challenging conditions and environments. We have not seen any adverse interaction between our system and other competing systems either in the Bay Area or Pittsburgh — and we don’t expect to.”
While there haven’t been conflicts between competitive “self-driving” cars in the Iron City, or anywhere else, there have been a few notable mishaps with driven vehicles and pedestrians that can be related to the systems programming. And not all the participants survived.
One thing everyone seems to agree upon is that safety must come first. “Automakers must agree to stop competing on safety,” says the NSC’s Hersman. “Everyone wants to buy a top safety pick. But that only says one car is better than another. Automakers need to take a page out of the aviation industry’s playbook. They need to voluntarily elevate safety to their primary focus and share their research, what they have learned through experimentation to ensure that not only one automaker will make the safest [AVs], but all will be the safest.”
While sharing data — any data — is a hot topic right now that no one wants to discuss or be associated with thanks to Facebook and Cambridge Analytica, some believe it is the lynchpin to the success of the self-driving movement going forward. “Sharing work is essential to developing robust technologies that can communicate and work in tandem over a standard network that has not yet been defined,” says Bryan Reimer. “It is the only way to produce the safest, best-driving AI possible.”
Edge cases — rare events that tax the capabilities of autonomous systems – might be one of the most compelling reasons for automakers to share knowledge. Think of other drivers unexpectedly swerving, debris on the road, or plastic bags blowing in front of a vehicle. Because such events occur infrequently, and computers currently lack the common sense to decide how to respond, training AVs to cope with edge cases is hard.
Most automakers are not giving up their “competitive advantage” easily – safety procedures are intellectual property to them.
But by sharing with each other information from edge cases that have happened, AV firms can test their systems in simulators to see how they would respond, and adjust them where needed, benefiting from each other’s experience.
“We need to have some minimum performance standards or expectations for creating a ‘belt and suspenders’ approach toward safety,” says NSC’s Hersman.
The question is what kind of data should be shared. Many automakers are actively talking to regulatory bodies and other members of the automotive community to discuss best practices. But most are not giving up their “competitive advantage” easily – safety procedures are intellectual property to them.
“Not all data should be shared in all circumstances,” says Rami Sass, CEO of Whitesource, which provides software development and security teams full control and visibility over their open source usage. “But data that will have an impact on safety features and on the ability of AV to maintain safety and security will have to be a common effort for [the transition from driven to driverless] to work properly.”
While compatible code is not necessarily a concern to most, everyone agrees that a common language to facilitate communication between vehicles is a must. Even so, vehicle-to-vehicle communication is not a must-have element in the self-driving equation.
“It simply becomes another sensor, gathering information on the state of traffic, the position of other vehicles, and their speed,” says Danny Shapiro, director of Automotive for chip-set giant Nvidia. The chipset behemoth has taken a dominant position over the past few years in the development of super-fast in-vehicle computers that allow cars to drive autonomously. “That data will help cars see around corners sooner, identify oncoming traffic sooner, tell us to adjust speed to avoid a collision,” says Shapiro. “But there needs to be a common language, something all systems can understand.”
Unfortunately, V2V is a long way from being standardized. “So, we are now building systems that can make stand-alone decisions, not connected to another car or by connecting to the cloud but by being able to perceive their surroundings,” says Shapiro.
Right now, the goal is to have cars that can drive themselves safely within a certain geographic area within the next few years. They will be used for mobility-as-a-service applications. Ford, GM, Tesla, Uber, and Waymo have all promised to launch autonomous ride-share operations, and are well on their way to delivering on that promise.
However, these first generation of AVs won’t be able to drive themselves anywhere, anytime, under any conditions. For that to happen most experts believe communication between vehicles — hence, compatibility — is necessary if the vehicles want to coexist without conflict. So, when will this happen? Your guess is as good as anyone’s. If autonomous ride shares are a commercial success, developers might not have incentive to take that next step: a fully autonomous car in your driveway.