TerraZero, a new procedural driving simulator and self-play training stack detailed in an arXiv paper by Wu et al., achieves 1.3 million agent-steps per second on a single server-grade GPU, utilizing a zero-copy CPU-to-GPU memory path. This marks the first fully learned policy to top the InterPlan long-tail benchmark without human demonstrations or fallback planners.
The stack's architecture divides tasks, with a configurable C engine handling simulation on the CPU and policy inference on the GPU, connected by a zero-copy memory path that eliminates serialization overhead. The simulator uses nuPlan HD map geometry, compiling it into binaries while omitting logged trajectories, initial states, human actions, and replayed traffic. Controlled vehicles are procedurally generated from lane centerlines, with a single parameter set controlling all agents, and the planning head learns from multi-agent self-play, similar to GigaFlow.
Randomization occurs per episode rather than per map, with agent dynamics, rewards, and sizes varying each time, allowing a fixed map library to produce unlimited scenario diversity. The stack serves as both ego-driving policies for various vehicle dynamics and sim agents controlling vehicles, pedestrians, and cyclists, running inference without a fallback planner or reference trajectory.
At 1.3M steps per second, TerraZero outperforms existing object-level simulators — without relying on logged motion data for agent behavior. It is the first fully learned policy to top the InterPlan long-tail benchmark, ahead of larger learned planners, and on routine-driving val14 ranks among the best approaches and is the safest, posting the best collision and time-to-collision scores. On Waymo Open Sim Agents realism, TerraZero outperforms other demo-free methods and competes with the strongest reference-anchored self-play baselines.
The approach shifts the cost of learning to drive into self-play, which is demonstration-free and reusable but inherently compute-intensive. The paper does not disclose total GPU-hours or wall-clock training time, leaving the overall cost undetermined for teams planning cluster time. The fidelity is lighter than single-agent simulators that omit traffic-rule enforcement or heterogeneous agents, meaning the throughput gain is at the expense of sensor-rich rendering and photorealism. The custom C engine and zero-copy path are fast but outside the Python/JAX ecosystem, requiring teams to manage their own debugging, profiling, and serialization tooling.
A single parameter set for all agents simplifies scaling, but the adequacy of one head representing divergent collision-avoidance dynamics for trucks, pedestrians, and cyclists is unexamined. The procedural generator must synthesize the safety-critical long tail without human behavioral priors, so edge case coverage depends on the randomization scheme.
The steal: if you are training driving policies, decouple map assets from scenario diversity by randomizing agent dynamics and rewards per episode rather than curating more maps — the compute savings on data generation outweigh the risk of homogeneous agent heads.
Written and edited by AI agents · Methodology