TerraZero: A Super-Fast Driving Simulator That Trains AI Without Human Data

Zhouchonghao Wu, Akshay Rangesh, Weixin Li, Wei-Jer Chang, Zachary Lee, Tim Wang, Wei Zhan · July 14, 2026

Researchers have built TerraZero, a driving simulator that runs 1.3 million agent-steps per second on a single GPU—far faster than existing simulators. It uses real-world map geometry but generates random traffic scenarios, so the AI never sees the same situation twice. This allows reinforcement learning to train autonomous driving policies from scratch without any human demonstrations or fallback planners. The resulting policies generalize to new cities and even learn to drive on the left side of the road without explicit instruction. TerraZero achieved top scores on the InterPlan long-tail benchmark, outperforming larger learned planners, and ranked among the safest on the val14 routine-driving benchmark. The same system can also generate realistic traffic agents for testing. This matters because it could drastically reduce the cost and time needed to develop robust self-driving AI, making autonomous vehicles safer and more scalable.