Enabling Realistic Simulations for Autonomous Vehicles
An open consortium to advance human-centric simulations for safer autonomous vehicles.
Project Overview
Updated October 23, 2024.
The Problem
Autonomous vehicles (AVs) hold the promise of forever changing how we use cars, but before driverless cars become mainstream, we need to ensure they are safe. However, experts estimate it could take billions of kilometers and hundreds of years of driving to ensure AVs are fully reliable. AV and advanced driver assistance systems (ADAS) developers have turned to computer simulations to train and test driverless cars. The realism of these simulations is critical, as differences between the simulated driving environment and the real world can introduce unacceptable safety and performance errors, requiring far longer and more expensive development cycles. This “simulation-to-real gap” is a major hurdle for AV development, particularly for interactions between the AI vehicle and other simulated humans, such as other drivers, pedestrians and cyclists.
Human behaviour is highly varied, and the inability to adequately represent these nuances in simulation can prevent autonomous or semi-autonomous vehicles from handling diverse situations on the road. Jurisdictions currently experimenting with driverless cars on public roads have seen these real-world errors firsthand, where AVs in dynamic, complex situations can exhibit a variety of safety and performance issues. AV software stacks to address these issues are currently a patchwork of open-source tools, commercial products and platforms, internal resources and digital asset libraries – lacking a truly accessible and interoperable solution for large-scale simulations of human drivers and other road participants’ behaviour.
How We Are Solving It
Led by Inverted AI in collaboration with Embodied AI Foundation (CARLA), this project aims to establish a consortium to accelerate the technological and commercial development of behavioral models for AV/ADAS simulation and related fields.
The consortium intends to unite innovators from across the sector to develop a comprehensive product and partner ecosystem, enhancing the accessibility, interoperability, and realism of simulation technology, while creating and capturing new value in the global market. In its first phase, the project will scale Inverted AI’s AV simulation software and incorporate it into CARLA’s platform to create a large-scale solution that can integrate with other platforms and tools to overcome obstacles with simulation testing in the AV/ADAS industry. The consortium will use cloud infrastructure providers’ resources and AI compute to make the solution available to AV developers, vehicle manufacturers, and academic researchers around the world.
By bringing together diverse partners in industry and academia, this project will make human simulation software for AV industry more efficient, cost-effective, and realistic — accelerating development of autonomous vehicles worldwide and bolstering Canada’s reputation as an AI development leader.
