Model Predictive Supervisory Control for Hierarchical and Distributed UAS Traffic Management

What the math gives to ML

The paper contributes a constructive supervisory-control mechanism: represent safety, workflow, resource exclusivity, and task constraints as deterministic automata, then combine a safety supervisor with receding-horizon cost optimization. The transferable asset is an exact discrete action mask generated from the current automaton state, together with a short-horizon planner that selects among only admissible event sequences. In neural networks, this can become a runtime shield for RL policies, sequence decoders, or learned multi-agent planners, preserving hard safety under uncertainty and distribution shift. The strongest testable claim is binary: if the automaton model is correct and every action passes through the supervisor, forbidden-event violations should be exactly zero.