Edge-Enabled Hybrid Deep Learning for Real-TimeMulti-Agent Tactical Maneuver Prediction

Sammanfattning

Autonomous navigation for unmanned ground vehicles remains difficult in complex outdoor terrain, especially when local sensing is limited and useful information may also come from cooperating aerial agents. This thesis addresses that problem by developing a cooperative UAV–UGV simulation framework in which a Husky UGV navigates through the Baylands environment with support from two UAV scout agents. The aim is not only to compare trajectory-prediction models offline, but also to examine how learned predictions behave when they are returned to the simulation workflow for inference-based evaluation. The work begins with a rule-based cooperative maneuvering system that provides safe mission execution, recovery behavior, and interpretable reference actions. This rule-based controller is then used to generate supervised learning data through synchronized simulation, rosbag recording,episode-frame export, and sliding-window sample construction. From eight recorded runs, the final dataset contains 51,818 ordered frame records and 51,706 supervised trajectory samples. Four learned models are trained and compared under the same data split: CNN–LSTM, CNN–GNN–LSTM, CNN–GNN–Transformer, and CNN–GNN–LSTM–Transformer. Offline evaluation is performed using average displacement error (ADE), final displacement error (FDE), and root mean square error (RMSE). All learned models improve over the constant-velocity baseline, and the CNN–GNN–Transformer achieves the strongest offline result with ADE 0.0094, FDE 0.0060, and RMSE 0.0144. The selected model is then deployed back into the simulation using fixed trained weights. It completes the navigation task, showing that the learned trajectory predictor can support end-to-endinference evaluation, although the rule-based controller remains more stable and computationally lighter. The framework is also extended with communication-aware relay profiles that emulate latency, jitter, and packet loss in the cooperative UAV-to-UGVinformation link. These tests show that degraded communication affects mission duration and cooperative context quality, while the system remains operational under the tested profiles. The thesis contributes a complete teacher-to-student workflow for cooperative predictive navigation, a practical method for converting rule-based robot missions into trajectory-learningdata, and an experimental basis for studying learned UAV–UGV cooperation under imperfect communication.