A 100-agent swarm syncs on 18.2% of the channel: weaker batteries broadcast less, and the 1 m clearance still holds

Swarm synchronization traffic can shrink with each node's battery instead of exchanging state on a fixed cycle. Standard periodic consensus broadcasts on a fixed cadence even when states barely change, making radio traffic the dominant energy cost on battery-powered edge nodes. In a simulation of N = 100 dynamic agents over T = 1000 synchronization cycles, ET-S3 accumulated 3.35 MB against 18.40 MB for periodic consensus and 6.35 MB for fixed-threshold event-triggered control.

The safety constraint sits inside the event-triggered consensus loop rather than being encouraged indirectly through reward shaping. Decentralized particle swarm optimization and MADDPG bias toward safe behavior only in expectation and still cross the 1.0 m minimum spacing during dense convergence. Under the same simulated setting, ET-S3 reported a 0.0% safety violation rate, decentralized particle swarm optimization 12.3%, and MADDPG 3.2%.

Each agent broadcasts only when its own state-estimation error exceeds a threshold formed by a time-decaying term plus an energy penalty that grows as the battery drains. Fixed-threshold event-triggered control compares error magnitude alone and does not account for how much charge a node has left. The energy penalty is a positive constant divided by remaining energy, so the threshold rises as charge falls and low-battery nodes skip small state oscillations.

Violation risk is corrected by an online quadratic program that keeps the control input as close as possible to the consensus input while satisfying the barrier constraint. An uncorrected consensus input can drive an agent outside the safe set. When the barrier condition would be violated, the quadratic program replaces the nominal input with the minimum change that keeps the safe set forward invariant.

The results rest on a co-simulation using Python, PyTorch 2.3, Ray RLlib and NS-3 with N = 100 dynamic agents over T = 1000 cycles, with edge hardware modeled as ARM Cortex-M4 class nodes and LoRaWAN and Wi-Fi 6 low-power radios. The authors propose hardware-in-the-loop tests on physical Cortex-M4 nodes next, to validate battery-level energy savings and whether quadratic-program execution times stay within control-loop deadlines.

The authors note this is a simulation study and battery-level savings await validation on physical nodes; the barrier-function guarantee relies on correct neighbor state information, and stale estimates or heavy packet loss require a larger safety margin; the scalability sweep stopped at N = 500, and the effect of highly heterogeneous radio links on trigger timing remains open.