# Robot steering-angle prediction lightweight network based non-local attention and lane guidance

## TL;DR

This paper introduces a new lightweight network for predicting robot steering angles, combining attention mechanisms and lane guidance to improve accuracy and speed.

## Contribution

A novel end-to-end network integrating non-local attention and lane guidance for efficient and accurate robot steering prediction.

## Key findings

- The proposed model achieves a 54.88% increase in inference speed compared to the baseline.
- The model reduces MAE and RMSE by 8.47% and 18.23%, respectively.
- Combining Non-Local Block and Ghost Module improves model expressiveness and efficiency.

## Abstract

Predicting the steering angle of robots is a core challenge in autonomous navigation. This paper proposes a novel end-to-end prediction network that integrates non-local attention and lane line guidance mechanisms to significantly reduce computational time and improve prediction accuracy. Built upon the ResNet architecture, the network incorporates a Non Local Block mechanism to enhance global context modeling and a Ghost Module to reduce parameter count and improve feature extraction efficiency. To further optimize training, a ReduceLROnPlateau learning rate scheduler is employed to adaptively adjust the learning rate, effectively mitigating overfitting. Additionally, a lane line annotation method, which combines Canny edge detection with the Hough transform, is used to semantically guide the input images. This enhances the representational power and generalizability of the training data. Experimental results demonstrate that the proposed network outperforms the baseline across multiple evaluation metrics. Under identical experimental conditions, the proposed model achieves a 54.88% increase in inference speed and reduces the mean absolute error (MAE) and root mean square error (RMSE) by 8.47% and 18.23% respectively. Ablation studies further confirm that the combination of the Non-Local Block and Ghost Module significantly improves both expressive capacity and computational efficiency of the model. These findings suggest that the proposed method offers a high-precision, efficient, and low-latency visual perception solution for real-time autonomous navigation of wheeled robots in complex environments.