There’s at least one thing policymakers can quickly agree on: for autonomous vehicles to become mainstream, they must be proven safe.
Policy, tech and automotive leaders gathered this week at NVIDIA’s GPU Technology Conference in Washington to discuss how the industry can safely and effectively deploy self-driving technology. While companies and regulators are still determining the best practices in many areas of the autonomous frontier, one point was certain: the development process must be transparent and openly communicated to build trust with consumers.
“Trust is essential to adopting new technologies. It requires open, honest communication and matching words with actions,” Heidi King, deputy administrator of the National Highway Traffic Safety Administration, said during her address at GTC. “Transparency is necessary to build public trust.”
King pointed to NVIDIA’s own Self-Driving Safety Report, issued earlier this week, as a good example of companies communicating with the public, providing resources to educate on self-driving technologies and processes.
On a panel I moderated that included speakers from the auto industry, Department of Transportation and Congress, collaboration throughout the industry was highlighted as the best way to communicate with and educate the public.
“Education is part of this joint project of autonomous driving, where government and industry work together,” said Brad Stertz, director of Government Affairs at Audi.
In addition to education, speakers identified processes that can be used now to build public trust and ensure the safe deployment of autonomous driving in the future.
“These are great opportunities for us to learn from each other about what’s working and what’s not working,” said Bert Kaufman, head of Corporate and Regulatory Affairs at Zoox.
Simulation: ‘A Key Element’
Before manufacturers introduce self-driving cars to public roads, they must rigorously test and validate the technology so that it will operate as intended on its own. In the past decade, the industry has used measures such as test miles driven and the number of disengagements — any time a safety driver must take back control of the vehicle — as benchmarks for validation. However, regulators don’t see those as comprehensive enough.
“We’ve been using the concept of miles traveled as a way to measure progress, but we’re realizing it’s not sufficient,” said Derek Kan, undersecretary of Transportation for Policy at the Department of Transportation during a fireside chat. “Simulation likely plays a key element.”
To address these gaps in validation, NVIDIA has developed an AV simulator called NVIDIA DRIVE Constellation. This data center platform enables manufacturers to test autonomous driving hardware and software at scale before going on the road.
Running NVIDIA DRIVE Sim software, DRIVE Constellation can generate a synthetic, detailed driving environment or use real-world sensor data to test how an autonomous vehicle would react to specific driving scenarios. The platform contains the same AI car computer that would run in the vehicle, NVIDIA DRIVE AGX Pegasus, allowing manufacturers to measure how the entire hardware and software stack operates.
With this solution, cars can experience rare and dangerous traffic conditions, test different sensor arrangements and retest real-world situations over and over again. It’s a much more rigorous approach than counting miles driven or disengagements.
Simulation isn’t the only development area where diversity and redundancy are necessary. To build a technology that the public can truly trust, manufacturers must build that robustness into every step of the process.
“Certainty is important,” Kan said. “If a person does not feel certain, they will not feel safe in an autonomous vehicle. The onus is on the industry to really provide a clear framework on how it works.”
An autonomous vehicle has thousands of moving parts and millions of lines of code. To provide the certainty Kan and other policymakers are calling for, each of those parts needs a backup, using a different process to provide the same function.
To do so, NVIDIA uses diverse and redundant deep neural networks and algorithms for autonomous driving functions, and builds diversity and redundancy into the hardware, with many different processor types built into a single system-on-a-chip.
For example, one DNN in the car is dedicated to detecting signals such as stop signs or traffic lights to determine when to stop. Another DNN also works to identify stop conditions, but relies on the context of the scene — crosswalks or heavy traffic — ensuring the vehicle will know to stop in variety of wait conditions. Dealing with variations in the environment around a self-driving vehicle is critically important and is the reason why we don’t hard-code the software.
By incorporating these safety practices into the development process now and working together for the future, NVIDIA and policymakers are laying the groundwork for a new transportation era that will truly transform the way we live, work and play. But most importantly, it will make sure that future is safe.
The most important feature of autonomous driving isn’t AI, but the safety it enables.