Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.13028v1

Read the source

Source

Plain-Language Brief

TerraZero is a procedural driving simulator and training stack for reinforcement learning at scale, designed to train autonomous driving agents without human demonstrations or fallback planners. The simulator provides realistic scenarios by generating randomized rule-based road users, signal controllers, and dynamics, rewards, and sizes per episode. It is faster than existing object-level simulators while keeping fidelity lighter than single-agent systems.

What the Document Says

  • The paper presents TerraZero, a procedural driving simulator for reinforcement learning at scale.
  • TerraZero runs simulation on the CPU and policy inference on the GPU over a zero-copy path.
  • It sustains 1.3M agent-steps per second on a single server-grade GPU.
  • The simulator omits features like heterogeneous agents, multiple dynamics models, and full traffic-rule enforcement compared to single-agent systems.
  • TerraZero uses logged data for real-world map geometry only.
  • It populates each map with randomized rule-based road users and signal controllers.
  • Agent dynamics, rewards, and sizes are randomized per episode.
  • Policies are trained from scratch by reinforcement learning alone on a compute-efficient self-play recipe across GPUs.
  • TerraZero does not use human demonstrations or fallback planners at inference.
  • Policies generalize zero-shot across cities and datasets, including emergent left-hand-traffic driving without explicit supervision.
  • TerraZero is the first fully learned policy to top the InterPlan long-tail benchmark.

Key Concepts Mentioned

  • Autonomous driving
  • Simulator
  • Reinforcement learning
  • Scale
  • Procedural generation
  • Zero-demonstration self-play
  • Performance evaluation benchmarks

Claims Made by the Document

  • TerraZero is faster than existing object-level simulators.
  • TerraZero keeps fidelity lighter than single-agent systems.
  • Policies trained with TerraZero generalize zero-shot across cities and datasets.

Evidence Used

  • 1.3M agent-steps per second on a single server-grade GPU
  • Comparison with single-agent systems

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

  • The authors do not discuss potential challenges or limitations of their approach.

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

  • How does TerraZero compare to other simulators in terms of realism and robustness?
  • What are the potential challenges or limitations of training autonomous driving agents with TerraZero?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper does not mention the specific GPU used for testing.
  • It is unclear how TerraZero compares to other simulators in terms of realism and robustness beyond the mentioned metrics.

Raw Abstract / Excerpt

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.