TerraZero: Procedural Driving Simulator for Scalable Self-Play

What changed

TerraZero is a novel procedural driving simulator and self-play training stack developed to address the limitations of existing autonomous driving simulation platforms. Traditional simulators often face a trade-off between simulation speed and fidelity, making them too slow for the extensive reinforcement learning (RL) required for robust agent training. TerraZero aims to bridge this gap by combining high-speed simulation with realistic scenario generation. The system utilizes a configurable C engine for simulation on the CPU and policy inference on the GPU, connected via a zero-copy path. This architecture sustains up to 1.3 million agent-steps per second on a single server-grade GPU, significantly outperforming many existing object-level simulators. Unlike lighter systems, TerraZero maintains fidelity by supporting heterogeneous agents (vehicles, pedestrians, cyclists), multiple dynamics models, and full traffic-rule enforcement.

Key to TerraZero's approach is its procedural scenario generation. Instead of relying solely on logged real-world data, which often lacks rare but critical situations, TerraZero uses logged data primarily for its map geometry. Each map is then populated with randomized, rule-based road users (including reactive vehicles, pedestrians, and cyclists), randomized traffic signal controllers, and per-episode randomization of agent dynamics, rewards, and sizes. This proceduralization ensures that a single map can yield an effectively unbounded set of diverse training scenarios, covering the safety-critical long tail of driving situations.

Furthermore, TerraZero employs a compute-efficient self-play training recipe. Policies are trained from scratch using RL alone, with zero human demonstrations, imitation learning, or fallback planners at inference. This approach allows the resulting policies to generalize zero-shot across different cities and datasets, even demonstrating emergent left-hand-traffic driving without explicit supervision. The system has been validated on multiple datasets, including Waymo Open Motion Dataset, nuPlan, and CARLA, showing competitive performance on benchmarks like InterPlan and Waymo Open Sim Agents Challenge.