Researchers present TerraZero, a procedural driving simulator and self-play training stack designed for training robust autonomous driving agents without human demonstrations. The system utilizes a configurable C engine to run simulation on the CPU and policy inference on the GPU over a zero-copy path, sustaining 1.3M agent-steps per second on a single server-grade GPU.
- Policies train from scratch via reinforcement learning with zero human demonstrations and no fallback planner at inference.
- The simulator treats logged data only as map geometry, populating maps with randomized rule-based road users to create unbounded scenarios.
- As an ego policy, TerraZero tops the InterPlan long-tail benchmark and ranks among the safest approaches on routine-driving val14.
- On Waymo Open Sim Agents realism, the method outperforms other demonstration-free methods and is competitive with reference-anchored self-play.
TerraZero allows policies to generalize zero-shot across cities and datasets, including emergent left-hand-traffic driving, while serving as both a driving policy simulator and an agent controller for vehicles, pedestrians, and cyclists.