CARIL: Confidence-Aware Regression in Imitation Learning for Autonomous Driving

Elahe Delavari¹ Aws Khalil¹ , Jaerock Kwon^1
1University of Michigan-Dearborn

🎥 Demo Video

📄 Abstract

End-to-end vision-based imitation learning has demonstrated promising results in autonomous driving by learning control commands directly from expert demonstrations. However, traditional approaches rely on either regressionbased models, which provide precise control but lack confidence estimation, or classification-based models, which offer confidence scores but suffer from reduced precision due to discretization. This limitation makes it challenging to quantify the reliability of predicted actions and apply corrections when necessary. In this work, we introduce a dual-head neural network architecture that integrates both regression and classification heads to improve decision reliability in imitation learning. The regression head predicts continuous driving actions, while the classification head estimates confidence, enabling a correction mechanism that adjusts actions in low-confidence scenarios, enhancing driving stability. We evaluate our approach in a closed-loop setting within the CARLA simulator, demonstrating its ability to detect uncertain actions, estimate confidence, and apply real-time corrections. Experimental results show that our method reduces lane deviation and improves trajectory accuracy by up to 50%, outperforming conventional regression-only models. These findings highlight the potential of classification-guided confidence estimation in enhancing the robustness of vision-based imitation learning for autonomous driving.

🖼️ Method Overview

📊 Evaluation Results

Classification Performance

The classification head achieved an overall accuracy of 76.33%. The highest precision was observed in class 5 (straight driving), reflecting its dominance in the dataset. However, the model struggled more with underrepresented classes (e.g., sharp turns). Misclassifications primarily occurred between similar steering classes.

Real-Time Confidence Estimation

The model estimates the confidence of its predictions using classification entropy and maximum softmax probability. When the model is confident (left image), the regression prediction is retained. In low-confidence situations (right image), a correction is applied. This enhances safety by dynamically adapting predictions in uncertain scenarios.

Route Similarity Analysis

Confidence-guided correction significantly improved trajectory accuracy, particularly in complex scenarios (e.g., two-turn routes). Results show a reduction in deviation of up to 50% across multiple metrics including Fréchet distance, DTW, Area Between Curves (ABC), and Curve Length (CL).

Fréchet & DTW Metrics

ABC & Curve Length

Overall Route Similarity

📖 Citation

  @article{delavari2025carilconfidenceawareregressionimitation,
    title={CARIL: Confidence-Aware Regression in Imitation Learning for Autonomous Driving}, 
    author={Elahe Delavari and Aws Khalil and Jaerock Kwon},
    year={2025},
    eprint={2503.00783},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2503.00783}, 
}