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

TL;DR

The paper tackles the lack of action-level confidence in end-to-end imitation learning for autonomous driving by proposing a dual-head network that jointly performs continuous regression and discrete action classification to estimate confidence. This confidence output enables a correction mechanism, applying uncertainty-informed adjustments to steering in real time. Evaluated in the CARLA simulator, the approach improves trajectory accuracy and stability, with up to a 50% reduction in route deviation compared to regression-only baselines, and demonstrates practical, real-time confidence estimation. The work enhances robustness and interpretability in vision-based imitation learning, offering a public implementation and suggesting pathways to further improve uncertainty handling and integration with reinforcement learning.

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. The source code is available at https://github.com/ElaheDlv/Confidence_Aware_IL.

Figures (5)

• Figure 1: Overview of the proposed dual-head architecture for vision-based imitation learning. The model takes a front-facing camera image as input, processed by an encoder ($\varepsilon$) to extract latent features. Two separate heads generate outputs: the classification head ($C$) estimates confidence levels by predicting probability distributions over actions, while the regression head ($R$) predicts continuous control values such as steering and acceleration. The classification output enables real-time confidence assessment, allowing corrective actions when necessary, improving the reliability of imitation learning-based autonomous driving. (Icons are from Flaticon.com)
• Figure 2: Sample images of the RGB front camera.
• Figure 3: Routes in CARLA's Town04 used for trajectory evaluation. Route A (green) represents a one-turn path, Route B (blue) is a straight segment, and Route C (red) is a two-turn route.
• Figure 4: Confusion Matrix for the Classification Model. The $x$ axis is for the predicted classes and the $y$ axis is for true classes.
• Figure 5: Real-time confidence estimation of the regression steering value. (Left) The vehicle remains within the lane, resulting in high confidence. (Right) As the vehicle deviates from the lane, the confidence decreases, indicating higher uncertainty in the regression output. video available: https://www.youtube.com/watch?v=2RPf-T_lsLc