Recreating Diverse Traffic Scenarios for Benchmarking Embodied AI Safety in Simulation 

Abstract: 

Autonomous driving (AD) as an emerging technology is revolutionizing mobility and transportation. Despite the enormous progress, it remains challenging to evaluate driving systems being developed directly in the real world due to AI safety concerns. To tackle this, driving simulation platform becomes a stepping stone for testing AD systems before their real-world deployment. In this case, diverse and realistic traffic scenarios that reflect the real-world complexity are crucial for evaluating the AI safety in simulation. I will introduce our effort of building the MetaDrive driving simulator with massive number of synthetic traffic scenarios. MetaDrive can import real-world vehicle trajectories and learn to generate novel ones. It can substantially improve the realism and diversity of the traffic scenarios in simulation as well as thorougly evaluate and improve the decision-making of AD systems.