Deep Learning, Safety, and the Industrialization of Autonomous Mobility

Karl Iagnemma and Oscar Beijbom from Aptiv Autonomous Mobility discuss the evolution, challenges, and future of deep learning in self-driving cars. Karl, President of Aptiv Autonomous Mobility, introduces Aptiv's background (spun from Delphi, formerly GM) and its global scale as a Tier 1 automotive supplier focusing on safer, greener, and more connected solutions. He recounts the rapid progression of autonomous vehicle technology from the DARPA Urban Challenge in 2007 to Aptiv's current operations, including a significant deployment in Las Vegas where their autonomous cars, integrated with the Lyft network, have provided over 30,000 rides to 50,000+ passengers, driving over a million miles with a 4.95-star customer rating. This operational experience highlights the daily complexities autonomous vehicles must navigate, from avoiding obstacles and managing unpredictable road users to handling construction zones. A central argument presented by Karl is the community's shift away from the initial belief that a single, end-to-end "black box" deep learning architecture could directly translate pixels to actuator commands. This shift is primarily driven by the paramount concern of safety, which he dissects into two axes: technical safety (provable, validated performance within an operational design domain) and perceived safety (rider confidence and trust). The challenge lies in building trust in neural networks for safety-critical systems, requiring rigorous validation to ensure they meet user and regulatory requirements. Karl outlines three critical dimensions for validating neural network-based systems: trusting the data, trusting the implementation, and trusting the algorithm. Trusting the data involves ensuring its completeness, representativeness, lack of bias, and accurate annotation. Trusting the implementation addresses the reliability of hardware and code execution. Trusting the algorithm focuses on its performance in rare, critical events, its invariance and stability to small perturbations (e.g., adversarial images), and its interpretability—the ability to explain why a specific decision was made, especially in the event of a crash. He also highlights challenges like the statistical difficulty of proving safety for rare events (e.g., 275 million miles without a crash), evolving regulatory requirements, and the complexities of revalidation for continuous code updates or new operational environments. Despite these challenges, Karl emphasizes that deep learning methods are incredibly useful when embedded within a broader safety argument. He advocates for "caging the learning" by surrounding powerful, neural network-based sub-modules (like trajectory proposers) with more rigorous, provably safe systems. This approach, leveraging concepts like functional safety and SOTIF, allows for a convincing holistic safety argument. Oscar Beijbom then elaborates on practical applications, tracing the deep learning revolution from image classification (AlexNet) to object detection (R-CNN, SSD) and presenting Aptiv's work on a LiDAR pipeline using "point pillars" and the "nuScenes" benchmark dataset, demonstrating high-performance 3D bounding box generation at high inference speeds without temporal smoothing.

I'm not allowed to talk about anything technical or interesting I'm gonna flout that a little bit and and raise some topics that we think about that I think are interesting you know questions too to keep in the back of your mind as you're thinking about deep learning an autonomous driving.

we as primarily scientists and technologists didn't have a great idea how we're gonna industrialize this technology and actually bring it to market and make it reliable and robust and make it safe.

our star rating today is four point nine five which is pretty good key numbers we've given this point over 30,000 rides to more than 50,000 passengers we've driven over a million miles in Vegas and a little bit additional but primarily there and as I mentioned the 4.95.

there was a period a few years ago where I think as a community people thought that we would be able to go from pixels to actuator commands with a single learned architecture a single black box I'll say generally speaking we no longer believe that's true.

this safety dimension is tricky because there's actually there's actually two axes here one is the actual technical safety of the system... but there's another dimension which appear is you know called perceived safety.

It really comes down to this question of how do we convince ourselves that we can validate these systems again validating the system ensuring that it can it can meet the requirements the operation requirements in the domain of interests that are imposed by the user.

if you believe rant and they're pretty smart folks they say you got to drive 275 million miles without accident without a crashed you can claim a lower fatality rate than a human with 95% confidence.

this question of the properties of the algorithm its invariance its stability is something of high interests and then lastly and one more point to this this notion of interpretability.

we tend not to think about using a neural network as an holistic system to drive a car but we'll think about it as a sub-module that we can build other systems around.

engineering is inching closer to the Natural Sciences... We're creating things that we don't fully understand and then we're investigating the properties or creation.