Data Scaling Laws for Imitation Learning-Based End-To-End Autonomous Driving
Yupeng Zheng, Pengxuan Yang, zhongpu xia, Qichao Zhang, Yuhang Zheng, Ben Lu, Teng Zhang, Chao Han, Weize Li, Songen Gu, Xianpeng Lang, Xiangyuan Lan, Dongbin Zhao
AI summary
Problem
End-to-end autonomous driving lacks comprehensive scaling law analysis due to limited real-world data, leaving the impact of data quantity versus distribution on real-world generalization unclear.
Approach
The authors compiled a million-scale real-world driving dataset (ONE-Drive) and trained imitation learning models at varying data scales, evaluating them through both open-loop trajectory fitting and closed-loop simulation.
Key results
- Open-loop trajectory fitting follows a power-law scaling relationship with data volume
- Closed-loop simulation performance does not follow a power law, exposing a metric mismatch
- Targeted augmentation of long-tailed scenarios yields disproportionate performance gains
- Scaled data enables combinatorial generalization to novel scenes and actions
Why it matters
Provides actionable guidelines for scaling training data and prioritizing scenario diversity to safely deploy end-to-end autonomous driving systems.
Abstract
The end-to-end autonomous driving paradigm has recently attracted lots of attention due to its scalability. However, existing methods are constrained by the limited scale of real-world data, which hinders a comprehensive exploration of the scaling laws associated with end-to-end autonomous driving. To address this issue, we collected substantial data from various driving scenarios and behaviors and conducted an extensive study on the scaling laws of existing imita- tion learning-based end-to-end autonomous driving paradigms. Specifically, approximately 4 million demonstrations from 23 different scenario types were gathered, amounting to over 30,000 hours of driving demonstrations. We performed open- loop evaluations and closed-loop simulation evaluations in 1,400 diverse driving demonstrations (1,300 for open-loop and 100 for closed-loop) under stringent assessment conditions. Through experimental analysis, we discovered that (1) the performance of the driving model exhibits a power-law relationship with the amount of data, but this is not the case in closed-loop evaluation. The inconsistency between the two assessments shifts our focus toward the distribution of data rather than merely expanding its volume. (2) A small increase in the quantity of long- tailed data can significantly improve the performance for the corresponding scenarios; (3) appropriate scaling of data enables the model to achieve combinatorial generalization in novel scenes and actions. Our results highlight the critical role of data scaling in improving the generalizability of models across 2026 IEEE International Conference on Robotics and Automation (ICRA 2026) June 1-5, 2026. Vienna, Austria 979-8-3315-8160-2/26/$31.00 ©2026 IEEE 22151