TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale

TerraZero, a novel procedural driving simulator, has been developed to facilitate the training of robust autonomous driving agents through reinforcement learning at scale. This simulator is designed to be fast, realistic, and diverse, allowing it to cover the long tail of safety-critical scenarios that are rarely encountered in logged data. By utilizing a configurable C engine, TerraZero's simulation capabilities can be tailored to specific training needs. The simulator's ability to generate diverse and realistic scenarios enables autonomous driving agents to learn from a wide range of experiences, improving their overall robustness and reliability. TerraZero's implications extend beyond the realm of autonomous driving, as advancements in reinforcement learning can have significant effects on policy, security, and workforce dynamics¹. As such, the development of TerraZero has significant implications for the future of autonomous systems, making it an important area of study for practitioners and researchers alike.