Gatik unveils Arena, a next generation simulation platform for autonomous trucks

“One of the things that the industry has struggled with for a while is that there have been novel approaches that came about a few years ago that allowed re-creation or more photorealistic synthesis of data or synthesis of sensors,” said Apeksha Kumavat, co-founder and chief engineer at Gatik, in an interview with FreightWaves. “However, the grounding in physics or the physics-inspired way of doing these things was lacking, which caused skepticism in terms of using this for real safety validations.”

Traditional simulation methods relied on game engine physics that produced unrealistic sensor data, limiting their usefulness to small segments of the autonomous vehicle stack. Arena overcomes these limitations by providing end-to-end stack simulation with physically accurate sensor data.

The integration with Nvidia’s Cosmos world foundation model significantly expands Arena’s capabilities. While Gatik’s proprietary tools handle 3D reconstruction and scenario generation, Cosmos enables the translation of scenarios to different environments. One example is creating an environment in Europe compared to the U.S. Other possibilities include regional variations, like types of buildings, infrastructure and intersections.

This technology allows Gatik to test trucks in conditions they might rarely encounter physically, such as snow in Texas. Other examples include adverse weather conditions, unpredictable road users, and complex urban interactions.

“What Cosmos essentially provides is a physics-grounded, more generative AI style that allows large-scale simulation volumes,” Kumavat said. “As we collect the data, we collect a few miles on those networks, and as we plug that into the pipeline of Arena, we are able to turn those few miles to a few thousand and millions of miles.”

In the past, these were situations that would be too expensive, time-consuming, or unsafe to test in the real world.

For autonomous vehicle companies, the addition of these virtual environments is critical, as it would take years and millions of miles before some of the edge cases would be found using traditional testing.