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

Training autonomous driving agents is challenging due to the need for fast, realistic, and diverse simulators. Current simulators often lack the speed or realism required for effective reinforcement learning. TerraZero addresses this by providing a procedural simulator that generates diverse driving scenarios and trains policies from scratch without human input. Builders might find this approach valuable as it allows for scalable training of robust driving agents across various environments.

Introduces a novel procedural driving simulator that enables scalable reinforcement learning.

Demonstrates strong performance across multiple benchmarks with rigorous evaluation.