3CIL: Causality-Inspired Contrastive Conditional Imitation Learning for Autonomous Driving

Huanghui Zhang, Zhi Zheng, Xiaomin Lin

Keywords: Imitation Learning, Autonomous Driving, Causal Reasoning, Causal Confusion

Abstract: Imitation learning (IL) aims to recover an expert's strategy by performing supervised learning on the demonstration datasets. Incorporating IL in safety-crucial tasks like autonomous driving is promising as it requires less interaction with the actual environment than reinforcement learning approaches. However, the robustness of IL methods is often questioned, as phenomena like causal confusion occur frequently and hinder it from practical use. In this paper, we conduct causal reasoning to investigate the crucial requirements for the ideal imitation generalization performance. With insights derived from modeled causalities, we propose causality-inspired contrastive conditional imitation learning (3CIL), a conditional imitation learning method equipped with contrastive learning and action residual prediction tasks, regularizing the imitator in causal and anti-causal directions. To mitigate the divergence with experts in unfamiliar scenarios, 3CIL introduces a sample-weighting term that transforms the prediction error into an emphasis on critical samples. Extensive experiments in the CARLA simulator show the proposed method significantly improves the driving capabilities of models.

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9150