AFD-CAMLs: Agile Force-Distribution-Aware Planning and Control for Cable-Suspended Aerial Multi-Lifting Systems

Can a planner and controller explicitly select and regulate cable-force distributions in redundant cable-suspended multi-UAV lifting systems without sacrificing agile payload tracking?

With four or more UAVs, the same payload wrench can be produced by many cable-tension combinations, so accurate payload-pose tracking does not guarantee a safe load distribution. Geometric mismatch, cable-length errors, and low-level tracking errors can shift the realized tensions, overloading some UAVs or letting cables go slack, especially when the tension-to-wrench allocation matrix is near degenerate.

Earlier work includes quasi-static position-based formation methods, cascaded wrench-allocation force controllers, and more recent centralized whole-body NMPC trajectory planners. Quasi-static methods ignore dynamic coupling and do not guarantee feasible tensions; force-based methods are limited by time-scale separation and sensitivity to payload parameter uncertainty; the trajectory-based NMPC method improved agility but did not directly plan or regulate the internal force distribution.

The framework adds three components to a centralized NMPC transport stack. A global planner explores the null space of the payload wrench-allocation matrix to produce tension-feasible cable-force and direction references under a specified outward preload. The NMPC stage and terminal costs are augmented with cable-direction and force-reference terms. An admittance filter compares planned tension with an onboard force-observer estimate and applies virtual mass-spring-damper corrections to each UAV's position, velocity, and acceleration references.

In simulation, a four-UAV level hover had an allocation inverse condition number of 0.0012, and activating the admittance filter substantially reduced tension error; an inclined case at 0.1916 remained low even without filtering. A ten-UAV simulation showed the same decentralized correction scaled and reduced tension imbalance. With a 1 cm cable-length mismatch in agile simulation, the filter reduced 3D payload mean absolute error from 0.221 m to 0.199 m. In real four-UAV hover, tensions were strongly imbalanced before filter activation and converged toward references after activation, though a residual offset of about 2 N remained versus about 0.5 N in simulation. In real agile flight, the filter improved 3D payload tracking error from 0.1787 m to 0.1525 m; with an unmodeled 0.6 kg sloshing basketball load, tracking was 0.1974 m. Aggregate cable-tension RMSE fell from 3.51 N to 2.09 N.

The authors report a residual tension offset of about 2 N in real hover, likely due to unmodeled aerodynamic effects such as rotor downwash on the payload that the onboard observer does not capture. They plan future work on online tuning of admittance parameters from the allocation-matrix inverse condition number and on preserving NMPC constraints under admittance corrections. The evaluation is limited to a Vicon-equipped indoor arena, four real UAVs and up to ten simulated UAVs, with no code release or peer-reviewed venue noted.