Reproduce “RealEngine: Simulating Autonomous Driving in Realistic Context” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junzhe Jiang, Nan Song, Jingyu Li, Xiatian Zhu, Li Zhang

RealEngine is a simulation platform for autonomous driving that integrates 3D scene reconstruction with dynamic traffic participant composition, achieving photorealistic multi-modal sensing and closed-loop interaction. It demonstrates only a 1% gap in driving agent performance between simulated and real-world sensor inputs and outperforms existing simulation benchmarks in visual fidelity and generalization.

Key insights

  • Decoupling static background scene modeling from dynamic foreground traffic participant composition allows for both high-fidelity neural rendering of real-world environments and flexible, interactive simulation of agents.
  • Sophisticated pose and appearance correction methods for scene reconstruction, combined with video generative priors, are crucial for enabling photorealistic novel view synthesis that supports free-form trajectories.
  • A comprehensive closed-loop evaluation metric, incorporating penalty and weighted reward terms, effectively reveals limitations of AD agents (e.g., safety critical failures) not apparent in open-loop assessments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.