Reproduce “Bench2Drive-R: Turning Real World Data into Reactive Closed-Loop Autonomous Driving Benchmark by Generative Model” 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
Junqi You, Xiaosong Jia, Zhiyuan Zhang, Yutao Zhu, Junchi Yan
Researchers at Shanghai Jiao Tong University developed Bench2Drive-R, a generative framework that converts real-world driving data into a reactive closed-loop simulation environment for end-to-end autonomous driving (E2E-AD) evaluation. This framework generates authentic, high-fidelity, and temporally consistent sensor data, achieving a Frechet Inception Distance (FID) of 10.95 on the nuScenes validation set and demonstrating improved perception and planning metrics compared to existing generative models.
Key insights
- Decoupling sensor rendering from behavioral rollout in a generative framework allows for high-fidelity, interactive simulation that maintains spatiotemporal consistency over long horizons.
- Noise modulation applied to previous frames during training, combined with Gaussian blurring, is critical for mitigating cumulative errors and achieving stable autoregressive generation in interactive simulations.
- A retrieval-based scene-level control mechanism, enhanced with 3D positional encoding, hierarchical sampling, and Reference Classifier-Free Guidance, effectively leverages real-world background fidelity while improving robustness to inference deviations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.