International Conferences
| Event-Aware Relabeling for Addressing Future Leakage in End-to-End Autonomous Driving |
| Year |
2026 |
| Month |
|
| Journal |
2026 / International Conference on Artificial Intelligence in Information and Communication (ICAIIC) |
| Author |
Haechul Chang, Siewoo Kim, Adeeb M. Islam, Seong-Jun Kim, and Seung-Hyun Kong* |
|
Recent advances in autonomous driving have shifted the field from modular pipelines to End-to-End (E2E) frame- works. E2E frameworks directly map raw sensor inputs to control commands, leveraging large-scale data to learn robust driving policies. However, the conventional data generation process relies on future actions and trajectory information as Ground Truth (GT). This procedure introduces a “Future Leakage” problem, where information unavailable at inference time unintentionally enters the training process. This leads to causal confusion, where the model learns spurious correlations instead of true causality, thus compromising driving safety. In this paper, we propose an Event-Aware Relabeling technique to address this problem. The proposed method detects interaction events with obstacles and vehicles by utilizing traffic lights and the ego-vehicle’s state information. Then it effectively eliminates the leakage of future information by performing appropriate relabeling on the data prior to the event occurrence. Experimental results demonstrate that our method effectively mitigates the causal confusion problem and achieves a driving score improvement of approximately 14% compared to baseline. This study highlights that a high-quality data curation process is essential to ensure the safety of E2E autonomous driving systems and validates the potential of data-centric performance optimization. |
