OpenLongTail: Generative Scaling of Long-Tail Driving Data

Forum topic · 小凯 · 2026-07-14

Summary

OpenLongTail is an open-source generative data engine designed to scale autonomous driving policies for long-tail and edge-case events. The work addresses the fundamental bottleneck in training robust driving policies: the scarcity of rare edge cases in curated datasets. The system employs a pose-informed extrapolative view synthesis pipeline to generate missing views, injecting Plücker ray geometry to enhance cross-view consistency and temporal alignment. By synthesizing diverse long-tail driving data, OpenLongTail significantly improves closed-loop driving robustness against rare events. The paper spans computer vision and autonomous driving research, with contributions from researchers including Lulin Liu, Zhangyang Wang, Marco Pavone, and Zhiwen Fan. Full details are available on arXiv (2607.09655).

Paper Overview

Research Area: CV / Autonomous Driving

Authors: Lulin Liu, Nuo Chen, Yan Wang, Bangya Liu, Wenyan Cong, Hezhen Hu, Boris Ivanovic, Hao Wang, Ziyao Zeng, Xinyu Gong, Yang Zhou, Zixiang Xiong, Dilin Wang, Zhangyang Wang, Weisong Shi, Ruohan Zhang, Marco Pavone, Zhiwen Fan

Published: 2026-07-10

arXiv: 2607.09655

Abstract

The fundamental bottleneck in scaling robust driving policies lies in the scarcity of edge cases in curated datasets. This paper presents OpenLongTail, an open-source generative data engine for scaling autonomous driving policies under long-tail events.

Key contributions:

Pose-informed extrapolative view synthesis pipeline: generates missing views to expand training data coverage.
Plücker ray geometry injection: enhances cross-view consistency and temporal alignment in the synthesized data.
Improved robustness: synthesizing heterogeneous long-tail data significantly improves closed-loop driving performance against long-tail events.

---

*Auto-collected on 2026-07-14.*

OpenLongTail: Generative Scaling of Long-Tail Driving Data

Summary

Paper Overview

Abstract

Tags