17 April 2024 11:00 - 11:30
Generating realistic human behaviors in AV simulations
Simulation is a key component of validating the safety of autonomous vehicles, gradually replacing expensive on-road tests.
The main limitation to its applicability is the domain gap between the real world and the simulated one, and numerous advances in both classical simulation methods and generative AI are gradually reducing this gap in terms of dynamics, visuals, sensor physics and content.
In this talk, Adam will outline Inverted AI’s efforts in bridging the remaining behavioral gap, that is how the actions of simulated humans differ from the actions of real humans. He will discuss the key difficulties faced by existing approaches and how to build a better solution, including where to find suitable data, what architectures to use, how to connect them to existing workflows, and what it means for the future of AV development.