Title: StandardE2E: A Unified Framework for End-to-End Autonomous Driving Datasets

Abstract

The autonomous driving landscape is undergoing a significant transition, moving away from traditional modular pipelines that separate perception, prediction, and planning. Instead, the field is increasingly adopting end-to-end (E2E) models that translate sensor data directly into vehicle control signals, frequently incorporating auxiliary objectives like 3D detection, motion forecasting, and HD-map perception. While this advancement is fueled by a rapidly expanding array of sensor-dense driving datasets, the ecosystem currently faces fragmentation. Each dataset utilizes distinct file formats, application programming interfaces (APIs), coordinate systems, and modalities. Consequently, researchers are forced to rein preprocess data and conduct cross-dataset experiments on a project-by-project basis.

To address these challenges, we introduce StandardE2E, a framework designed to offer a singular, unified interface for E2E driving datasets. StandardE2E achieves this through three primary capabilities: (i) it consolidates dataset-specific preprocessing into a common data schema; (ii) it enables the integration of multiple datasets within a single PyTorch DataLoader, facilitating cross-dataset pretraining, auxiliary-task supervision, and scenario-level filtering; and (iii) it simplifies the integration of new datasets to a single mapping function that converts raw frames into the canonical schema, thereby keeping the rest of the downstream pipeline intact. The framework currently supports six datasets natively: Waymo End-to-End, Waymo Perception, Argoverse 2 Sensor, Argoverse 2 LiDAR, NAVSIM (OpenScene-v1.1), and WayveScenes101. The project is released as the open-source standard-e2e Python package, which can be accessed at https://github.com/stepankonev/StandardE2E.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC