This study focuses on RGB image classification across six common outdoor ground textures for mobile robots. It aims to develop a lightweight network that achieves high classification accuracy while maintaining low computational cost and low inference latency on embedded hardware. The proposed OGTM-Net employs a compact feature-extraction architecture for ground-texture classification, constructs a 960-dimensional feature representation, and introduces a 1D batch normalization layer before the classifier. Under the same experimental settings with five common optimizers, OGTM-Net achieves relatively high test accuracy among the six evaluated architectures. On the NVIDIA Jetson Orin NX embedded platform, the results demonstrate that OGTM-Net can achieve high-performance ground-texture classification while reducing computational complexity and maintaining low measured inference latency. These characteristics provide an additional lightweight network option for outdoor ground-texture classification on embedded platforms for mobile robots.