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

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

arXiv: 2607.13028Published: July 14, 2026

AI summary

TL;DR

TerraZero is a procedural driving simulator for training autonomous driving agents using reinforcement learning. It achieves high simulation speeds and realistic scenarios without relying on human demonstrations. TerraZero's policies generalize well across cities and datasets.

Why It Matters

TerraZero's approach enables the development of more robust and generalizable autonomous driving agents, which is crucial for real-world applications. Its ability to simulate diverse scenarios and adapt to different environments can improve safety and efficiency in autonomous driving. This technology has the potential to accelerate the development of autonomous vehicles.

Key Findings

1.3M agent-steps per second simulation speed
zero-shot generalization across cities and datasets
state-of-the-art performance on InterPlan long-tail benchmark

Topics

reinforcement learning

autonomous driving

procedural simulation

self-play

zero-shot learning

Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.