Will This Tech Replace Real-World Safety Tests?

Yahoo

25-03-2025

Volvo is using AI to create life-like virtual environments to test its ADAS technology that used to require on-road testing in controlled environments.
The automaker is using Gaussian splatting to create realistic 3D environments and subjects sourced from real-world visual data to train its systems.
The technology, which was unavailable a decade ago, now promises to train ADAS and autonomous software quicker and more thoroughly by training it on "edge cases."
A decade ago on-road testing was seen as an irreplaceable part of the development of advanced driver assistance systems (ADAS), with automakers racking up tens of thousands of miles to fine-tune various features.
But slowly, the virtual world has been taking over this arena, especially when it comes to the development of higher levels of autonomous driving.
Volvo is now using AI to create life-like, virtual worlds to aid in the development of its safety software including ADAS features—something that wasn't available just a few years ago.
Specifically, the automaker is using a method called Gaussian splatting to create from scratch realistic 3D environments and subjects sourced from real-world visual data collected by cars.
What's more, these environments can be easily tweaked as needed, adding or changing the behavior and number of various road users and traffic patterns to test out how a safety system would react in various scenarios.
"We already have millions of data points of moments that never happened that we use to develop our software," says Alwin Bakkenes, Head of Global Software Engineering at Volvo Cars.
The footage of the environments themselves wouldn't fool a human, but it's realistic enough for camera-based ADAS to train themselves, and for developers to examine the various systems' reactions.
One of the advantages of generating realistic but artificial environments with minute variations is to discover dangerous and improbable edge cases—a term that has received plenty of attention in the autonomous vehicle sphere in recent years. Encountering such rare scenarios in traffic would ordinarily take tens of thousands of miles of driving data, but reproducing them in virtual worlds while adding small variations in each iteration helps Volvo train its safety systems.
The environments were created with help from Zenseact, an AI and software startup owned by Volvo.
"Thanks to Gaussian splatting we can select one of the rare corner cases and explode it into thousands of new variations of the scenario to train and validate our models against," Bakkenes adds. "This has the potential to unlock a scale that we've never had before and even to catch edge cases before they happen in the real world."
Still, there are limits to virtual environments, even very photorealistic ones, as more advanced ADAS now rely on multiple camera, lidar, and radar inputs simultaneously. So the amount of information going into ADAS decision-making is expanding as well.
If there is another limit to AI-created environments, it's that they are not speeding up the creation of a single regulatory jurisdiction for higher-level autonomous systems even in the US, even as they are speeding up and improving testing.
Of course, such techniques are not a 100% replacement for real-world, on-road testing. So Volvo and other automakers aren't relying entirely on 3D, AI-generated environments, even with a number of companies already specializing in creating such worlds for testing autonomous systems, including SAE Level 3 and Level 4 systems.
But the auto industry could be approaching a stage where the majority of time and miles are accumulated in completely artificial environments, with various systems being trained and validated on hundreds of thousands of miles of simulated highways and city streets.
Will we see a single, national standard emerge in the US for autonomous systems in the coming years, including SAE Level 3 that Volvo and Polestar want to offer, or will such systems remain governed on a state-by-state basis? Let us know what you think in the comments below.