EIDrive: an efficient and interpretable visual-language framework for low-level control signal prediction in autonomous driving

Abstract

In recent years, end-to-end autonomous driving models have shown promising potential by directly generating waypoints or control signals from sensor data. In particular, low-level control signals, such as steering angle and vehicle speed, directly determine vehicle motion, making their prediction accuracy and interpretability important to the safety and reliability of autonomous driving systems. However, existing methods mainly regress control signals from visual inputs and lack explicit semantic characterization of the basis underlying control generation, making it difficult to establish a traceable link between scene understanding and control outputs. To address this issue, we propose E fficient and I nterpretable Driving (EIDrive), an efficient visual-language framework for low-level control signal prediction in autonomous driving. Given monocular driving videos, EIDrive jointly performs control signal prediction and natural-language explanation generation. Specifically, a sparse orthogonal subspace projection mechanism is introduced to improve feature compactness and structural expressiveness. A lightweight fine-tuned BERT model is employed as a shared semantic backbone, together with a structured control prediction head to model structured correlations among control-related features. On the BDD-X dataset, EIDrive achieves a turning-angle RMSE of 6.1 and a speed prediction accuracy of 88.2% under the $$1.0\,\mathrm {m/s}$$ threshold. It also demonstrates competitive language-generation performance while maintaining a compact computational profile.