The auto industry is in the midst of a defining era as it works vigorously to safely increase the number of self-driving vehicles on public roads.

Waymo, which operates 600 self-driving vehicles in the Phoenix area, is an indication of how far the technology has evolved since 2009, when Google launched its self-driving car program, and since 2004, when teams of eager college students, entrepreneurs and corporations put early autonomous vehicles to the test in the first of three pivotal desert races staged by the Defense Advanced Research Projects Agency, or DARPA.

And yet, as the industry looks to build on that progress, it faces a number of challenges when it comes to creating a more robust environment for self-driving vehicles.

At a Shift panel discussion in June at Waymo headquarters in Mountain View, Calif., industry experts addressed the challenges that will impact the future of autonomous vehicles.

Simplifying the tech

One fundamental challenge is the complexity of the technology needed to develop and operate the vehicles.

A fully autonomous car requires an endless number of hardware and software components and algorithms, all focused on enabling an AV company such as Waymo to be able to see a self-driving vehicle’s surroundings and understand what is happening around the car, before charting a vehicle’s course from point A to point B.

“If you think about it, a lot of the challenges we faced back then are still the same problems that we are working on now,” said Mike Montemerlo, senior staff software engineer at Waymo.

Montemerlo, who was the software lead for the Stanford team that won the 2005 DARPA Grand Challenge, added, “It’s just that the level of complexity has increased dramatically.”

Still, most in the industry tend to agree: The advancement of self-driving vehicles will depend heavily on how well companies continue to find innovative ways to seamlessly integrate complexities into the vehicles.

“When we kicked this [self-driving car] project off in 2009, what we wanted to do is better understand the complexity of the problem and understand what it would take for us to go from a DARPA challenge type environment, that is very synthetic and artificial, to real public roads and interacting with real people,” said Dmitri Dolgov, Waymo chief technology officer. Like Montemerlo, Dolgov has been with Waymo since the company was known as the Google self-driving car project.

More efficient sensors

Finding new ways to improve the efficiency and capabilities of lidar, radar and camera sensor technology will continue to be crucial to create a more robust environment for self-driving vehicles.

Reengineering the Chrysler Pacifica into a Waymo AV entailed hundreds, if not thousands, of design changes to the sensors and computers to make them more capable and reliable, Waymo engineers said. Future autonomous sensor development will need to focus on more capable longer-range sensors, with more energy-efficient chips, experts said.

Smarter tools

In addition to more efficient sensors, many in the industry see deep learning — a machine’s ability to continually analyze and adapt to new situations — as a key to improving autonomous vehicle systems.

The technology is said to be more efficient than a traditional engineering approach, when it comes to helping AVs adapt to real-world driving dynamics.

Experts said further deep learning development will help improve self-driving vehicles’ performance in unpredictable scenarios. Take, for example, an object flying off the bed of a truck and landing in an AV’s path.

“We’re not just trying to detect pedestrians, but we need to detect distracted pedestrians and every object out there,” said Danny Shapiro, senior director of automotive at Nvidia, a Waymo partner whose hardware and software help enable AI-powered self-driving.

The need for more advanced mapping technology will continue. But Montemerlo said mapping will remain more of a support system for self-driving sensor technology.

“We came to the realization there’s this kind of sweet spot in the middle where you can use a map to help fill in the gaps in your perception in a way that you don’t believe it as absolute truth,” Montemerlo said.

More talent

The growing need to develop smarter technology across all sectors of the industry is driving an effort to find new ways to attract designers, engineers and developers.

“We are all chasing the same talent right now. And it’s a real struggle, not only to hire the right talent, but then to keep them on board,” said Christophe Marnat, executive vice president of the Electronics and Advanced Driver Assist Systems division, with ZF Group, a Waymo partner.

‘Chicken and egg problem’

Another big hurdle for self-driving vehicles is vehicle-to-infrastructure communication. Take, for example, the ability for autonomous vehicles to be able to communicate with smart traffic light systems, which helps to create a more robust environment for AVs.

“Some of the traffic lights are 30 years old … so there is a very low common denominator that you can rely on everybody having,” Montemerlo said.

“We’ll take advantage of that infrastructure when it becomes available. But you have a chicken and egg problem, where you can’t rely on it until it exists. And nobody wants to build it everywhere until it’s needed.”

Still, experts such as Marnat contend that developing more cohesive infrastructures for AVs is pivotal to creating a safer and more seamless self-driving experience for consumers.

“Functions like valet parking, for instance … can be heavily assisted by the investment in infrastructure,” Marnat said.

Fix a bug, unfix others

Testing and validation for self-driving vehicles will have a major impact on their advancement, experts contend.

“Back in the DARPA days we did do all kinds of testing,” Montemerlo said. “We spent almost three months living in the desert before the race. But our testing was still antidotal, that we would find a bug and then we would fix that bug. But we were never sure when we would fix that one bug that we weren’t unfixing three more.”

The advancement of AV technology will require companies to rely more on simulated data testing, which enables them to run thousands of variations of driving scenarios.

“There are two realizations that the industry has had over the past three to four to five years,” Shapiro said. “People were initially projecting, ‘Oh, self-driving cars in 2020, they’re going to be on the road.’ That was a little underestimated in terms of complexity, but the amount of infrastructure needed in a data center is mind-boggling now.”