Energy-Aware UAV-Aided MEC Network by Safe Reinforcement Learning
Abstract
Unmanned Aerial Vehicle (UAV)-aided Mobile Edge Computing (MEC) provides a flexible, rapidly deployable solution for extending coverage and low-latency computation in areas with limited terrestrial infrastructure. Achieving high performance requires the joint optimization of user association, CPU resource allocation, and UAV trajectory planning, since these interdependent decisions together determine connectivity, task execution time, and energy consumption. Dense multi-UAV deployments further demand strict collision avoidance, whereas existing approaches enforce safety only through soft penalties without formal guarantees. In this work, we investigate this joint optimization to minimize system energy consumption under constraints on UAV battery capacity, mobility, collision avoidance, and task deadlines. To solve it, we design MATD3-CBF-BC, which extends the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) framework with a Control Barrier Function (CBF)-based safety filter that enforces collision-free trajectories while avoiding energy-wasting detours. A Behavior Cloning (BC) term further guides each actor toward the CBF-corrected safe actions, which improves training stability. Extensive simulations show that MATD3-CBF-BC achieves lower energy consumption and shorter task completion time than benchmark algorithms while maintaining zero collisions.