---
title: "TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.13028v1"
permalink: "https://feed.astralor.com/items/itm_0f00e678420bdf439fa1989b"
published: "2026-07-14T17:59:02.000Z"
score: 79
channel: "ai"
tags: ["TerraZero","Reinforcement Learning","Procedural Driving Simulation","Autonomous Driving","Self-Play Training"]
---

# TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale

## AI commentary

> What excites me most about TerraZero is that pure self-play beats all supervised methods on the long-tail benchmark, offering hope for data-scarce extreme scenarios. For autonomous driving companies and researchers, if this simulator is open-sourced, it could change the training paradigm.

## At a glance

TerraZero is a high-efficiency procedural driving simulator designed for zero-demonstration self-play reinforcement learning. Its C engine runs simulation on CPU and policy inference on GPU via a zero-copy path, sustaining 1.3 million agent-steps per second on a single server-grade GPU, far faster than existing simulators. It uses real-world map geometry only, populating maps with randomized rule-based agents to generate unbounded scenarios, and trains policies from scratch without human demonstrations. The resulting policy tops the InterPlan long-tail benchmark among learned planners, achieves the best safety metrics on val14, and generalizes zero-shot to different traffic rules.

## Key points

- TerraZero achieves 1.3 million agent-steps per second on a single GPU, enabling extremely efficient training.
- Trains policies entirely via self-play RL with zero human demonstrations or expert data.
- Becomes the first fully learned policy to top the InterPlan long-tail benchmark with best safety metrics.
- Zero-shot generalizes to different cities and both left-hand and right-hand traffic rules.

## Full text

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 present TerraZero, a procedural driving simulator and self-play training stack. A configurable C engine runs 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, far faster than existing object-level simulators, while keeping fidelity lighter single-agent systems omit: heterogeneous agents, multiple dynamics models, and full traffic-rule enforcement. TerraZero treats logged data only as a source of real-world map geometry, populating each map with randomized rule-based road users and signal controllers and randomizing agent dynamics, rewards, and sizes per episode, so a map yields an unbounded set of scenarios. Every reported policy trains from scratch by reinforcement learning alone on a compute-efficient self-play recipe across GPUs, with zero human demonstrations and no fallback planner at inference. Policies generalize zero-shot across cities and datasets, including emergent left-hand-traffic driving without explicit supervision. As an ego policy, TerraZero is the first fully learned policy to top the InterPlan long-tail benchmark, ahead of larger learned planners; on routine-driving val14 it ranks among the best approaches and is the safest, posting the best collision and time-to-collision scores. On Waymo Open Sim Agents realism the same recipe outperforms other demonstration-free methods and is competitive with the strongest reference-anchored self-play method. One stack serves both roles: driving policies across dynamics for cars and trucks, and sim agents that jointly control vehicles, pedestrians, and cyclists.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_0f00e678420bdf439fa1989b