A team of researchers from Delft University of Technology and Wageningen University has shown that a concept borrowed from ecology, critical slowing down, can predict when a drone is about to lose control due to incremental damage, well before any visible signs of instability.
The study, published in PNAS, represents the first application of critical slowing down (CSD) indicators to actively controlled engineered systems with real-time feedback controllers. CSD is a generic dynamical phenomenon: as a system approaches a critical transition, its recovery rate from small perturbations slows progressively, manifesting as increased lag-1 autocorrelation (AC1) and increased variance in the system’s output signals. It has been previously documented in lake eutrophication, fisheries collapse, and abrupt climate shifts, but never before in an engineered system with a feedback controller.
The experiments
The team conducted 367 flights across two quadrotor platforms, the autonomous DragonFly (running an INDIFlight controller) and the human-piloted HoverFly (running Betaflight). They applied incremental asymmetric propeller blade tip damage from 0% (healthy) up to 55%, testing all four rotor positions across multiple flight conditions including hover, trajectory tracking, and windy conditions.
The key insight: even through a feedback controller continuously stabilizes the system, the combined closed-loop dynamics exhibit CSD as stability margins shrink due to damage. The indicator monitors observable signals, rotor speeds, accelerometer, gyroscope, without requiring an accurate model of the damaged system.
Results
The CSD indicators (lag-1 autocorrelation of rotor speed signals) increased monotonically with damage level across all rotors, including undamaged ones, well before visible instability occurred. At just 10% damage, the probability that AC1 was elevated above baseline was 0.59 (p<0.001). At 15% damage, it rose to 0.76 (p<0.001), with an effect size of 1.14 times the interquartile range.
The indicator also revealed asymmetric structural vulnerabilities that would not have been apparent from visual inspection: aft rotors showed significantly larger AC1 increases than front rotors (p<0.001), attributed to battery weight increasing demand on rear motors. Left rotor AC1 increases were larger than right (p=0.044 at 15% damage), traced to a manufacturing inconsistency where the aft left motor ran 5°C hotter.
The DragonFly (INDIFlight) lost control at approximately 30% blade damage, with clear AC1 elevation visible at 15%, well before failure. The HoverFly (Betaflight) remained stable up to approximately 55% damage. Under demanding conditions, wind combined with fast trajectories, loss of control occurred as early as 15% damage.
Broader implications
Because the approach is model-free and relies only on observable signals, the researchers argue it can be applied across diverse controlled systems: aircraft (the paper cites the Sriwijaya Air flight SJ-182 crash as a motivating example), industrial reactors, self-driving cars, power grids, and autonomous robots.
Two operational modes are proposed: real-time early warning during flight, and an exploratory “tinkering” mode where small perturbations are deliberately applied to empirically optimize system robustness during design.
Sources:
1. van Beers JJ, Scheffer M, Solanki P, van de Leemput IA, van Nes EH, de Visser CC. “Early warning signals for loss of control in complex systems.” PNAS. 2026;123(27):e2608847123. DOI: 10.1073/pnas.2608847123