Abstract
Unmanned aerial vehicles (UAVs) are increasingly deployed for delivery, monitoring, and emergency response, yet their large-scale coordination in congested airspace remains constrained by latency, bandwidth limits, and collision risks. Portable vertiports offer a mobile infrastructure for energy replenishment and structured takeoff/landing, but simultaneous operations require ultra-reliable and low-latency decision-making across multiple agents. This paper presents DREAM (DeepSetenhanced Reinforcement learning for Edge-based control in Ambiguous and dynamic environMents), a distributed edge intelligence framework that integrates deep reinforcement learning (DRL) with multi-access edge computing (MEC) principles. Each UAV operates as an edge client that performs real-time inference locally while synchronizing its policy with nearby vertiport servers through lightweight model updates, enabling scalable coordination without dependence on cloud backhaul. The framework adopts a centralized-training-and-decentralized-execution (CTDE) paradigm with a permutation-invariant DeepSet encoder and a safety-constrained Proximal Policy Optimization (PPOGAE) network. Simulation results under variable network latency and dynamic obstacles demonstrate that DREAM achieves a 98% mission success rate, reduces collisions by over 90%, and sustains stable performance within \mathbf{1 0 0 ~ m s} edge-to-UAV latency budgets. These results highlight the feasibility of edge-assisted multi-agent learning for autonomous vertiport operations and its potential integration into future 5G/6G-enabled aerial networks.