Volvo Cars is using AI-generated virtual environments to improve the development of its safety software, including advanced driver assistance systems (ADAS).
By synthesizing data from sensors in its latest vehicles – such as emergency braking, sharp steering and manual interventions – the company can reconstruct and analyze incidents in greater detail. This approach helps Volvo Cars better understand how to prevent similar situations in the future. The technology relies on a method called Gaussian splatting (below), which can generate realistic, high-resolution 3D scenes from real-world visual data.
In the virtual environment, Volvo Cars can add or remove road users and adjust the behavior of traffic or obstacles to create different scenarios. This enables the company to test its safety software across a wide range of traffic situations at a speed and scale that wasn’t previously possible. With this approach, Volvo Cars can develop software that performs well in complex, rare and potentially dangerous edge cases while reducing the time needed to test these situations from months to days.
Gaussian splat-synthesized swerve scenario
“We already have millions of data points of moments that never happened that we use to develop our software,” said Alwin Bakkenes, head of global software engineering at Volvo Cars. “Thanks to Gaussian splatting, we can select one of the rare edge cases and explode it into thousands of new variations of the scenario to train and validate our models against. 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.”
Volvo Cars combines virtual environments with real-world testing to train, develop and validate its safety software. These virtual environments are safe, scalable and cost-efficient. They are developed in-house in collaboration with Zenseact, an AI and software company founded by Volvo Cars. This work is part of a PhD program with leading Swedish universities, exploring the potential use of neural rendering techniques in future safety initiatives. The study is supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP).