One of the first sessions at CES 2025 included a panel discussion about how AI is learning to drive better—learning quickly and to drive fast—and touched on what will it take to safely and securely build, test, and deploy physical AI solutions for mobility.
With Paul Mitchell, Chairman and President, Indy Innovation Challenge, Inc. moderating, the discussion was titled "How To Build Physical AI for Mobility," (sponsored by Indy Innovation Challenge). Panelists included Sergio Matteo Savaresi, professor, Italian Trade Agency; Cynthia Kwon, SVP Product at Stack AV; and Vivek Moudgal, Vice President Strategy at dSPACE.
Savaresi kicked off the session, noting the real weakness of neural network approach to autonomous technology is the management of outliers, or edge cases, because 1% of these cases take 99% of the work that goes into safety and testing.
He noted, “AI is more an umbrella environment…where anything that can sense, act, reason and adapt can be classified as an AI. So if you look inside an application of AI for mobility, for instance, say, autonomous driving, you will discover that it's always a mix between model-based mathematics which is not considered AI—but it is AI—and black box, which is neural networks. The mix really depends on the application. Generally speaking, the use of neural networks is over-advertised, so in many cases, people claim they are using a lot of machine learning, a lot of neural networks, but if you look inside, they are using much more model-based algorithms because you can embed a lot of previous knowledge.”
Kwon added that her company employs deep neural networks, she said, which are “really important because when your system encounters a problem that it’s never seen before—a situation it’s never seen before—you need to be able to reason about what to do. However…sometimes, with these open-world models, or these deep nets, in a new situation, it might not give you the right output. It might not give you the right behavior because it has not seen this before, so in fact, part of what we do if we take a very structured approach and we make sure we build in plausibility checks, or functional safety checks, to ensure that the output from our deep net is actually the right behavior.”
Moudgal emphasized the importance of simulation in these environments, acknowledging that the ultimate test would be “if I would drive around and encounter every single scenario that the car would ever encounter in its lifetime, and I would have the best results, but that's not feasible. To be able to train, test, as well as validate the AI driver within a reasonable amount of time, as well as cost, makes simulations essential."
The Indy Autonomous Challenge (IAC) will take place on January 9 at the Las Vegas Motor Speedway from 2-4 PM PST, the Autonomous Challenge at CES 2025, offering attendees a firsthand look at the evolution of AI in high-speed motorsport.