Uncertainty-Aware Guard Conditions for Switching a Uav Swarm Into a Protective Mode Under Jamming and GNSS Denial

Swarms of unmanned aerial vehicles (UAVs) increasingly operate where the radio channel is jammed and satellite navigation is denied or falsified.In our earlier work [1] the swarm was controlled by a mode-conditioned multi-agent reinforcement learning (MARL) policy [2] supervised by a hybrid automaton [3] with three modes: Search, Track and Safe.The automaton decides which behaviour is active, and the learned policy decides how each UAV moves within it.The switch into Safe was triggered by a binary event: an agent is considered jammed when it lies inside a fixed jamming radius.Such a trigger describes the cause of the problem, not its consequence for the swarm, which is the loss of the ability to exchange information.Being a single threshold on an instantaneous value, it also makes a swarm that moves along the edge of the jammed zone switch modes back and forth, and it ignores navigation entirely, although a swarm without GNSS or with a spoofed GNSS fix is in danger even when every link works. Purpose and tasks of the study.The purpose of the study is to replace the binary event with guard conditions over statistics that the agents can actually measure, and to quantify how well each statistic recognises a critical situation.The tasks are: (1) to describe link loss, dead-reckoning drift and spoofing within one uncertainty model; (2) to define the swarm-level statistics and a decentralised way of computing them; (3) to build guards with hysteresis for entering and leaving Safe; (4) to evaluate the guards as detectors by Monte Carlo simulation, without a trained policy in the loop. Results and discussion.A packet from agent j to agent i is received with a probability that grows logistically with the signal-to-interference-plus-noise ratio at the receiver