Open X-AV: Unifying End-to-End Autonomous Driving Datasets

Abstract

The fragmentation of existing autonomous vehicle (AV) datasets hinders the development of generalizable driving policies that can handle complex and infrequent events. To overcome this, we introduce the Open-X AV (OXAV) repository, an initiative designed to aggregate a wide variety of AV datasets and enable models to learn from these diverse sources. We propose a two-stage training workflow using OXAV: a pre-training phase using perception-focused data, followed by post-training on challenging planning-centric scenarios. Our method DiffusionLTF, a simple end-to-end policy trained on OXAV, ranked second in the 2025 Waymo vision-based end-to-end driving challenge, demonstrating the benefits of diverse, aggregated data.

Publication
CVPRW 2025

Related