arxivcs.LGcs.AIcs.RO2026-07-14
TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale
Zhouchonghao Wu, Akshay Rangesh, Weixin Li, Wei-Jer Chang, Zachary Lee, Tim Wang, et al.
Training robust autonomous driving agents requires a simulator that is fast enough for reinforcement learning at scale, realistic enough to ground behavior in real-world map structure, and diverse enough to cover the safety-critical long tail that logged data rarely contains. We…