Transformer-Driven Visual Intelligence for Aerial Robotic Path Optimization

Mission-critical tasks like watching over sensitive areas, supporting disaster recovery efforts and supervising the environment require the use of UAVs which depend on good path planning for successful, safe and prompt performance. Planning methods for UAVs such as A*, Dijkstra and RRT, are largely used, but they perform better in still environments and not in areas full of obstacles. Thanks to CNNs, current learning-based methods are adaptable and handle local decisions based on images, though adapting to novel surroundings may not be so easy and usually requires significant re-training. To solve these problems, this study introduces a new transformer-based visual intelligence framework for real-time aerial robotic path optimization. By adding self-attention across various directions to a Vision Transformer, the model can deal with grid-based environment data and spot both general and specific features for UAV flight. Minimizing distance, power required and sharp changes in heading is included in the single optimization goal. Tests in simulation settings prove that the suggested approach reduces the path length by 16.5%, saves 27% of energy and runs 25% faster than both conventional planning and baseline CNN methods. The success achieved by the model is over 97% in urban, coastal, forest and mountainous environments. Incorporating spatial skills, perception of visual scenes and quick changes in flight, this investigation produces a strong and scaled answer for intelligent UAV route planning in automated navigation systems.