abstract
- This work presents an embedded system for real-time terrain surface classification with application to mobile robots. We employ MobileNetV2 and ResNet50 convolutional neural network models to distinguish between smooth, rocky, and muddy terrain. The networks were adapted and evaluated on a custom dataset, showing that MobileNetV2 offers a suitable balance between accuracy and efficiency for deployment on a Jetson Nano platform. As a use case, the classifier was integrated into an autonomous system, where the predicted terrain label dynamically selects the most appropriate trajectory tracking controller from a pre-designed control bank. This approach ensures that the control system adapts to the environment, enhancing the robustness of trajectory tracking. We present experimental results regarding classification accuracy, inference time, and a demonstration of the integrated closed-loop system. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.