A new research paper proposes using reinforcement learning to find the source of GPS jamming signals in environments where standard localization falls apart.
The approach treats the hunt for a radio frequency emitter as an active sensing problem. An agent moves through a space, collecting RF readings from a small 2x2 patch antenna, and builds up a picture of where the interference is coming from — one step at a time. Because a single measurement in a multipath-heavy indoor environment is almost meaningless on its own, the agent uses a recurrent policy to carry context across observations. The researchers tested two RL approaches — Deep Q-Networks and Proximal Policy Optimization — and ran experiments on simulated data generated with the Sionna ray-tracing tool, which models realistic signal propagation. The system hit an 80.1% localization success rate in those tests.
GNSS jamming is not a theoretical nuisance. It disrupts aviation, logistics, and emergency services, and indoor environments are among the hardest places to deal with it because walls and objects scatter signals unpredictably. An agent that can adaptively explore and reason across sequential readings is a more practical fit for that problem than a static sensor array.
The results come entirely from simulation, which the authors acknowledge — real-world multipath conditions will be messier than any ray-tracer produces. That gap between simulated success and field reliability is the standard caveat for this class of research, and 80.1% in a controlled virtual environment is a long way from a deployable tool.