Edge Intelligence for Autonomous UAV Missions: Scalable Trajectory Planning with AERPAW

Prasad Calyam

Professor, University of Missouri-Columbia

Prasad Calyam, a Professor at the University of Missouri–Columbia, and his students are advancing the way autonomous UAVs make real-time decisions in complex, resource-constrained environments. A key challenge in UAV mission planning lies in trajectory optimization: traditional centralized solvers, while powerful, scale poorly and are impractical for onboard use due to limited energy, compute capacity, and intermittent connectivity. This creates a critical gap for missions that require real-time, autonomous decision-making in the field.

To address this, Calyam and his team developed an edge-intelligent framework that combines Graph Attention Networks (GAT) with metaheuristic optimization. Their approach first uses GAT to identify and prune “low-utility” edges from the mission graph, effectively reducing the search space to only the most promising flight corridors. A Guided Local Search (GLS) algorithm then operates on this sparse subgraph to refine routes and identify near-optimal trajectories. By focusing computation on high-value options, the framework enables efficient path planning without relying on cloud-based solvers. Importantly, the entire pipeline is designed for execution on embedded hardware platforms such as NVIDIA Jetson, enabling fully autonomous, onboard decision-making.

AERPAW played a central role in shaping and validating this work, as the framework was explicitly developed for the AERPAW Autonomous Aerial Data Mule (AADM) Challenge, where UAVs must collect data from multiple ground stations and return within strict energy constraints. Using AERPAW Phase-1 deployment maps from the Lake Wheeler site, the team simulated realistic UAV-to-station communication networks and flight corridors, ensuring that their solution was grounded in real-world operational conditions. Through this integration, the research demonstrates how edge intelligence can transform UAV autonomy by enabling scalable, efficient, and practical trajectory planning in dynamic environments.

This work is reflected in recent and ongoing publications, including “Edge-Enabled Scalable Routing via Graph Neural Network Pruning and Metaheuristic Optimization,” presented at the ACM Workshop on Edge Intelligence in December 2025, and “Cross-Layer Deep Reinforcement Learning for Real-Time Multi-Drone Delivery from Mobile Depots under Uncertainty,” submitted to the IEEE ITS Journal in 2025.