UC-02 · Aviation use case

Separation Assurance in Shared Airspace

Keeping drones from multiple operators safely apart when fleets, policies, and navigation data differ, or are corrupted.

TC2GWUPurdueUT Austin
GPS observation corruption tolerated with near-zero near mid-air collisions by the robust policyGWU
~35%
fewer near mid-air collisions with runtime observation filtering under GNSS degradationGWU
~90%
decision accuracy of a fine-tuned language model acting as a deconfliction policyGWU
27% → 88%

The problem

In a busy city, several delivery companies may operate fleets with different aircraft, sensing ranges, and collision-avoidance logic in the same airspace. Every one of those tactical decisions depends on knowing where each aircraft is, and that knowledge comes from GPS, which can be degraded or spoofed.

Our approach

  • Mixed fleets, shared sky. GWU and Purdue examined whether independently trained deconfliction policies from different companies converge to safe co-existence in a simulated delivery network over Frisco, Texas, in the Dallas area, and whether the outcome is fair to operators with less capable aircraft. Policies converged and collisions fell sharply, but stronger fleets were systematically favored.
  • Robust to corrupted GPS. GWU and UT Austin model GPS corruption as a game between aircraft and a spoofer, derive a closed-form worst-case perturbation for efficient adversarial training, and add runtime safety filtering that corrects the observation the policy sees.
  • Language models as interpretable policies. GWU built a simulation-to-language pipeline in the BlueSky simulator and fine-tuned an open language model to make tactical deconfliction decisions, reducing near mid-air collisions and raising mission completion in closed-loop tests.

Looking ahead

Ongoing work addresses inference latency, formal safety assurance, and benchmarking against rule-based and reinforcement-learning baselines, with demonstrations planned for Year 3.

Figures

From the research

Two line charts comparing robust and nominal policies as observation contamination increases.
GWU · UT AustinSeparation assurance under increasing GPS corruption. The robust multi-agent policy keeps near mid-air collisions near zero and maintains larger separation than the nominal policy.
Pipeline from BlueSky scenario generation and prompt engineering to LoRA fine-tuning and closed-loop inference.
GWUSimulation-to-language pipeline for fine-tuning a large language model as a tactical deconfliction policy, then deploying it in closed-loop BlueSky simulations.
Training curves showing rewards converging for two companies and collisions falling while successes rise.
GWU · PurdueTwo delivery companies, each training its own attention-based policy, converge to co-existing deconfliction strategies in shared airspace.

GWU · PurdueUC-02Simulation

Two delivery fleets sharing the sky

BlueSky simulation of drones from two delivery companies over Frisco, Texas, each fleet flying its own attention-based multi-agent reinforcement-learning deconfliction policy.

GWUUC-02Simulation

A language model as a deconfliction policy

A fine-tuned large language model issues tactical deconfliction actions to drones in closed-loop BlueSky simulation.

Publications

  1. Fine-Tuning Large Language Models for Cooperative Tactical Deconfliction of Small Unmanned Aerial Systems

    I. Sharifi, A. Zongo, P. Wei

    CVPR Workshop on Multi-Agent Embodied Intelligent Systems Meet Generative-AI Era (MEIS), 2026

  2. Robust Multi-Agent Reinforcement Learning for Small UAS Separation Assurance under GPS Degradation and Spoofing

    A. Zongo, F. Fotiadis, U. Topcu, P. Wei

    IEEE Conference on Decision and Control (CDC), 2026Accepted

  3. Runtime Safety Filtering for Learned Small UAS Separation Policies under GNSS Degradation

    A. Zongo, P. Wei

    AIAA/IEEE Digital Avionics Systems Conference (DASC), 2026

  4. Separation Assurance between Heterogeneous Fleets of Small Unmanned Aerial Systems via Multi-Agent Reinforcement Learning

    I. Sharifi, H. T. Kim, M. H. Ahmed, M. Ghasemi, P. Wei

    IEEE International Conference on Intelligent Transportation Systems (ITSC), 2026