PaVSF: A Multimodal Fusion Power Consumption Prediction Model for Autonomous Wheeled Vehicles in Off‐Road Environments

ABSTRACT Accurate power consumption prediction is critical for ensuring long‐term autonomy of autonomous wheeled vehicles (AWVs) in unstructured off‐road environments. However, this task remains challenging due to the complex vehicle‐terrain interactions. Existing methods, predominantly relying on single‐modality data or pure physical models, fail to achieve the required prediction accuracy and cross‐domain generalization capability for practical deployment. To address these limitations, we propose Power‐aware Vision and State Fusion (PaVSF), a parameter‐aware multimodal fusion model. Our model conditions multi‐scale visual representations on vehicle operational states and adaptively fuses them with power signals. A physics consistency loss based on power conservation principles is integrated to regularize the learning process, enabling end‐to‐end power consumption regression. Experimental results demonstrate that our model significantly outperforms the self‐constructed BasicModel, reducing the MAE and RMSE of power consumption prediction by 76.16% (to 6.33 W) and 74.89% (to 8.68 W), respectively. Ablation study validates the necessity of each core component. Furthermore, the model exhibits exceptional generalization and robustness, as evidenced by low prediction errors in steady‐state intervals across real‐world unseen and heterogeneous terrains, with MAE values of 8.90 W on transition terrain in familiar states and 7.50 W on unmowed lawn terrain in unfamiliar states. This work presents a deployable solution for power consumption prediction in off‐road navigation. The code and related resources are available at: https://github.com/ShilongXu/PaVSF .