Researchers present TerraZero, a procedural driving simulator and self-play training stack designed for training robust autonomous driving agents without human demonstrations. The system runs simulation on the CPU and policy inference on the GPU over a zero-copy path, sustaining 1.3 million agent-steps per second on a single server-grade GPU.
- Uses logged data solely for real-world map geometry, populating maps with randomized rule-based road users and signal controllers.
- Randomizes agent dynamics, rewards, and sizes per episode to generate an unbounded set of scenarios from each map.
- Trains policies from scratch via reinforcement learning across GPUs with no fallback planner at inference.
- Achieves zero-shot generalization across cities and datasets, including emergent left-hand-traffic driving.
- Tops the InterPlan long-tail benchmark as the first fully learned policy to do so, while ranking among the safest approaches on routine-driving val14.
TerraZero serves as a unified stack for both driving policies across dynamics for cars and trucks, and sim agents that jointly control vehicles, pedestrians, and cyclists.