A Georgia Tech and MassRobotics team has proposed a four layer control stack, accepted to IROS 2026, the IEEE's flagship robotics conference, in which multiple lifting drones track and balance a cable suspended load using only their own cable measurements for load sharing, plus…
A team of delivery drones has a problem the moment more than one of them grabs the same load: someone has to be in charge. One aircraft has to know the payload's mass. Someone has to compute each rotor's share. A radio link has to stay up. Drop any of that, and the formation can rip the load apart or swing it into the ground.
GPAC, a four-layer control architecture proposed by Hadi Hajieghrary of Georgia Tech, Paul Schmitt of MassRobotics, and co-authors Benedikt Walter and Miguel Hurtado, sidesteps that entire coordination layer. Posted on arXiv as 2607.00024 and accepted to IROS 2026, the IEEE's flagship robotics conference, the design lets several quadrotors, the four-rotor drones most logistics research uses, carry a single cable-suspended payload without any of them acting as a lead. No drone is told the load's mass. No drone is told how many teammates it has. The only inter-agent communication is a low-rate neighbor-position broadcast used purely to avoid mid-air collisions. Everything else, including each drone's share of the total lift, emerges from local measurements of the cable attached to it.
That sounds like cheating. It is also a structural fact. A load shared by N aircraft obeys the physics of force balance. If a rotor reads the tension in its own cable and watches how that tension responds to its own thrust, it can solve backwards for the fraction of the load it is carrying. Call that estimate the local load share. Add the local load shares across all rotors, and the sum converges to the true total payload weight, even though no single aircraft knows the total or the headcount.
In a Drake simulation, a physics engine popular for robotics research, with flexible cables, sensor fusion, and a Dryden-style wind disturbance, a standard NASA turbulence model, this implicit scheme tracked the payload with a mean root-mean-square error of 33.8 centimeters and a 2.8 percent coefficient of variation across 13 random seeds. The same run produced exponential convergence of the load-share estimate, meaning the estimate approaches the true value at a rate that compounds rather than decays, and an input-to-state safety (ISSf) margin, a robustness bound on the controller, against a single activated constraint. Those numbers should be read on their own terms: they come from the authors' own simulation, not from outdoor flight tests, and the field-trial story has not yet been written.
GPAC stacks its behavior across four layers, each running at a different rate, as documented in the paper's Section II. The lowest layer runs a 50 Hz PID controller, the textbook workhorse of drone stabilization, supplemented here by an anti-swing term for the cable and an extended-state-observer feedforward that compensates for wind. Above it, a geometric attitude controller works directly on SO(3), the mathematical space of every possible 3D orientation. Layer 3 is a concurrent-learning estimator that updates its model of the payload mass in real time as new flight data arrives. Layer 4 wraps the stack in a safety filter built on a control barrier function (CBF), a constraint-enforcement routine that vetoes the controller if a safety condition is about to be violated. Each drone only needs an IMU, GPS, a load cell on its cable, and an encoder; mass identification happens inside the stack.
The interesting property of GPAC is what disappears from the design. A multi-drone lift normally demands a coordinator, a ground station or a lead aircraft, that knows the payload mass, plans every follower's trajectory, and disseminates load commands. That coordinator is also a single point of failure: drop the link, and the formation can tear the load apart or collide with it. By replacing those explicit messages with a physical invariant, GPAC trades a communication dependency for a structural one. The risk moves from the radio link to the cable itself, and the formation keeps working as drones are added or removed, because no part of the stack has to be re-tuned for a new team size.
That shift has limits. The 33.8 centimeter tracking figure is a simulated one. Real flight introduces actuator lag, GPS dropout, gusts that don't match a Dryden spectrum, and cable dynamics that the bead-chain model in the paper, a standard way to simulate a cable as a sequence of small rigid segments connected by springs, only approximates. The contribution today is a control architecture with provable convergence under stated assumptions, not a deployment-ready lift system. Replication on a real quadrotor fleet, ideally with hardware-in-the-loop wind, is the next obvious step.
The authors have cleared one administrative hurdle. Their preprint carries an IROS 2026 acceptance and a CC BY 4.0 license, with the arXiv submission trail dating to 20 June 2026. One co-author, Paul Schmitt, is affiliated with MassRobotics, the independent robotics industry group based in the Boston Seaport area. That tie is a signal that the work is being read inside an industry cluster, not a claim of industrial endorsement. The watch item is not the RMSE number. It is whether the architecture survives contact with real wind, a real cable, and a real payload, three things that, unlike the load share, can still surprise a drone.