Accuracy–Energy Trade-off Analysis of Lightweight Semantic Segmentation Models for Desert Off-Road Navigation

Autonomous navigation in desert environments poses great challenges due to the lack of structured roads, very similar terrain textures, and constantly varying illumination conditions. Semantic segmentation plays an important role in enabling robotic systems to understand terrain at the pixel level and make safer and more efficient navigation decisions. However, traditional evaluation of segmentation models is mainly focused on accuracy and inference speed and often neglects the effect of energy consumption, which is crucial for battery-powered edge devices operating in remote environments. This paper includes a comparative analysis of five lightweight semantic segmentation architectures, i.e., BiSeNetV2, MobileNetV3-Large, ResNet18-Seg, DDR-Light (Deep Dual-Resolution Network-Light), and SegFormer-B0, using a synthetic desert dataset generated via the Duality AI digital twin portal. The study tests each model on several parameters, such as mean Intersection over Union, inference speed, total energy use, carbon emissions estimates, and the overall efficiency score, which represents a trade-off between accuracy and energy use. Experimental results show that SegFormer-B0 achieves the highest segmentation accuracy, and MobileNetV3-Large achieves the best efficiency score because it has significantly lower power consumption. DDR-Light, even if competitive with respect to accuracy, has significantly higher energy requirements and is therefore less suitable for energy-constrained deployments. The results point to the fact that it is not enough to be accurate when deploying AI models on edge-based robotic systems, and energy-aware evaluation provides a more practical foundation for decision-making in real-world settings.