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

Zhouchonghao Wu, Akshay Rangesh, Weixin Li, Wei-Jer Chang, Zachary Lee et al.|July 14, 2026arXiv

Key Takeaway

You can train safer, more generalizable autonomous driving policies from scratch using only RL in a procedurally-generated simulator—no demonstrations needed, and it's fast enough to scale.

Summary

TerraZero is a fast procedural driving simulator that trains autonomous driving agents using only reinforcement learning, without human demonstrations. It runs 1.3M simulation steps per second on a single GPU while maintaining realism through traffic rules, diverse agent types, and randomized scenarios.

training agents

Key Terms

self-play procedural-generation zero-shot-generalization reinforcement-learning-from-ai-feedback