Dynamic-Aware Behavior Cloning with an Attention Mechanism for Lane-Keeping Assist on a Personal Mobility Vehicle

Abstract

Lane-keeping assist for low-speed personal mobility vehicles (PMVs) has to operate outdoors, where illumination changes, cast shadows and faint or partially missing lane markings degrade visual perception. End-to-end behavior cloning removes the need for hand-designed feature extraction, but the conventional formulation maps a single camera frame directly to a steering command and so ignores the vehicle state that determines how sharply the vehicle should turn. This paper presents Dynamic-Aware Behavior Cloning (DABC), a convolutional driving policy that takes the instantaneous linear velocity as an auxiliary input alongside the front-camera image, and that applies a learned gating branch to the convolutional feature map before regressing the angular velocity. A preprocessing and augmentation pipeline—region-of-interest cropping, YUV conversion, steering offsets for the side camera views, trimming of the straight-driving mode of the label distribution, random brightness and random shadow—is used to reduce the bias of the collected data. The policy was trained in a ROS 2/Gazebo simulation and then deployed on O-SEATER, a single-passenger differential-drive PMV. On an outdoor S-shaped track the trajectory RMSE was 0.1207 m at 0.2 m/s and 0.9945 m at 0.5 m/s. On a U-shaped track the RMSE ranged from 0.2567 m in the morning to 1.3480 m in the afternoon, when direct glare on the road surface obscured the lane markings. Robustness to strong and uneven illumination therefore remains the dominant limitation of camera-only lane keeping at this scale.