💻 computer science

Decentralized Geometric Control for Cable-Suspended Payload Transport with Adaptive Mass Estimation

This paper presents GPAC, a decentralized hierarchical control architecture that enables multiple quadrotors to cooperatively transport a cable-suspended payload without centralized coordination by utilizing implicit coordination for adaptive mass estimation, geometric control on nonlinear manifolds, and safety filters to ensure robust tracking and collision avoidance under wind disturbances.

Original authors: Hadi Hajieghrary, Benedikt Walter, Paul Schmitt, Miguel Hurtado

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hadi Hajieghrary, Benedikt Walter, Paul Schmitt, Miguel Hurtado

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of drones trying to carry a heavy, swinging chandelier through the air. Usually, you'd need a "team captain" drone to tell everyone exactly how hard to pull, or you'd need all the drones to constantly chat with each other about the chandelier's weight and position. If the captain gets knocked out or the chat line goes down, the whole team crashes.

This paper introduces a new system called GPAC (Geometric Payload Adaptive Control) that solves this problem. Instead of a captain or a constant chat, the drones work like a well-rehearsed dance troupe where everyone knows their part just by feeling the rope they are holding.

Here is how it works, broken down into simple concepts:

1. The "Silent Teammate" Strategy (No Captain Needed)

In most systems, every drone needs to know the total weight of the load and how many other drones are helping.

The GPAC Way: Each drone acts like a person lifting a heavy box with a rope. They don't need to know how many other people are helping or how heavy the box is. They just pull on their own rope.
The Magic: Because they all pull based on what their specific rope feels like, the math works out automatically. If you add more drones, they naturally share the load without anyone having to say, "Okay, I'll take 20%." They just pull until the rope feels right. This means if one drone drops out, the others just keep pulling; the system doesn't collapse.

2. Feeling the Wind and the Swing

Carrying a hanging load is tricky because it swings like a pendulum, and wind pushes it around.

The Swing: The drones have a special "anti-swing" mode. Imagine holding a bucket of water on a string; if you jerk it, the water sloshes. These drones are programmed to move smoothly so the "sloshing" (swing) dies down quickly.
The Wind: The system has a "sixth sense" (called an Extended State Observer) that guesses how hard the wind is pushing, even if the wind is invisible. It then pushes back against the wind automatically, keeping the load steady.

3. The "Safety Guard" (The Bouncer)

Even with a great plan, things can go wrong. The ropes might go slack (drooping), the drones might tilt too far, or they might get too close to each other.

The Bouncer: The system has a safety layer that acts like a bouncer at a club. It watches the drones constantly. If a drone starts to tilt too much or a rope gets too loose, the bouncer gently nudges the drone's controls to fix it before a crash happens.
Priority System: If two things go wrong at once (e.g., a rope is loose AND a drone is tilting), the bouncer has a rulebook: "Fix the loose rope first, then fix the tilt." This ensures the most dangerous problems are solved immediately.

4. Learning on the Fly (No Manual Weighing)

Usually, you have to weigh the cargo before you fly.

The Learner: These drones don't need to know the weight beforehand. As they lift, they use a "concurrent learning" method. It's like a student taking notes while studying. They look at how hard their motor is working and how the rope is pulling, and they quickly figure out, "Ah, I'm carrying about 1 kilogram." They do this so fast that even if the wind changes or the load shifts, they adjust their guess instantly.

5. The "Real-World" Test

The authors didn't just draw this on paper; they simulated it in a very realistic computer world.

The Simulation: They used a physics engine that mimics real wind, wobbly cables (like real ropes, not stiff sticks), and noisy sensors (like a GPS that sometimes glitches).
The Result: The drones successfully carried the load through a complex path (like a figure-eight). The load stayed within about 34 centimeters of its target path. That's roughly the length of a ruler, which is very precise for a swinging load in a windy simulation.

The Bottom Line

This paper presents a way for multiple drones to carry heavy, swinging loads together without needing a central boss, without needing to know the weight in advance, and without needing to talk constantly. They coordinate by "feeling" their own ropes and using a smart safety guard to prevent accidents. It's a system designed to be robust, safe, and able to keep working even if things get messy.

Technical Summary: Decentralized Geometric Control for Cable-Suspended Payload Transport with Adaptive Mass Estimation

Problem Statement
Cooperative aerial transport using multiple quadrotors and cable-suspended payloads offers increased capacity and fault tolerance compared to single-vehicle systems. However, these systems face significant challenges: nonlinear dynamics on manifolds (SE(3) × (S²)ᴺ), the need for decentralized operation without centralized coordination, and strict safety constraints (cable slack, excessive swing, collisions, and vehicle tilt). Existing solutions often rely on centralized state information, require knowledge of the total number of agents ( $N$ ) and payload mass ( $m_L$ ), or use linearized dynamics that lose stability during large-angle maneuvers. Furthermore, real-world factors like flexible cables, wind disturbances, and sensor noise are often neglected in theoretical frameworks. There is a lack of a unified framework that combines geometric manifold control, decentralized adaptive estimation without persistent excitation, and runtime safety supervision for multi-UAV cable transport.

Methodology: The GPAC Architecture
The authors propose GPAC (Geometric Payload Adaptive Control), a four-layer hierarchical architecture designed for $N$ quadrotors to transport a payload without a central coordinator or exchange of cable states/adaptive parameters. The system operates directly on the full nonlinear configuration manifold.

Core Insight (Implicit Coordination): Each quadrotor independently estimates its effective load share (payload mass share + its own cable mass) using only local cable measurements (tension and direction). The sum of these individual estimates naturally converges to the total weight the team must support, eliminating the need to know $N$ or $m_L$ in advance.
Layer 1: Position & Anti-Swing (50 Hz): Computes the desired force using PID feedback, feedforward compensation (gravity, trajectory, and cable tension), anti-swing control on the sphere ( $S^2$ ), and disturbance rejection. It reconstructs the payload position locally from the agent's own cable geometry.
Layer 2: Geometric Attitude Control (200 Hz): A geometric tracking controller on $SO(3)$ ensures almost-global exponential stability, avoiding Euler-angle singularities.
Layer 3: Adaptive Mass Estimation: Uses concurrent learning to estimate the payload mass share. This method utilizes stored data to achieve convergence without the "persistent excitation" typically required by adaptive laws, which is crucial during hover or slow maneuvers.
Layer 4: Extended State Observer (ESO): A third-order observer estimates lumped disturbances (e.g., wind) to provide feedforward compensation.
Supporting Modules:
- ESKF Sensor Fusion: A 15-state Extended Kalman Filter fuses IMU, GPS, and barometer data. All control loops close through this estimator, not ground truth.
- Priority-Ordered CBF Safety Filter: A modular layer supervises operational constraints (cable tautness, angle, swing rate, vehicle tilt, and inter-agent collision). It uses a sequential gradient projection approach rather than a full Quadratic Program (QP) to minimize computational cost. It employs Input-to-State Safety (ISSf) margins that scale with estimated disturbances.

Key Contributions

Coordinator-Free Control: A controller operating on the nonlinear manifold that requires no exchange of payload/cable states or adaptive parameters. Agents reconstruct the payload position locally, communicating only low-rate neighbor positions for collision avoidance.
Decentralized Adaptive Estimation: An estimator that converges exponentially to a uniformly ultimately bounded (UUB) neighborhood of the effective load share (including cable mass) using concurrent learning, even without persistent excitation.
Safety Supervision: A priority-ordered Control Barrier Function (CBF) layer that reduces risks of slack, excessive swing, and collisions. The paper proves that the force modifications induced by the safety filter keep the desired attitude within the almost-global stability region of the $SO(3)$ controller.
High-Fidelity Validation: A comprehensive simulation environment featuring flexible bead-chain cables, onboard sensor fusion, and Dryden wind turbulence, with all loops closed through the estimator.

Results
The system was validated in a high-fidelity Drake-based simulation with 3 quadrotors ( $N=3$ ) and a 3.0 kg payload under wind turbulence and asymmetric cable lengths.

Tracking Performance: The system achieved a mean 3D payload-tracking Root Mean Square Error (RMSE) of 33.8 cm (2.8% coefficient of variation over 13 seeds).
Estimation: The concurrent learning estimator converged to a steady-state mean of 1.18 kg per agent (true share 1.0 kg + cable mass), with bias decreasing as payload mass increased.
Safety & Constraints:
- Tension: Maintained positive tension in steady flight, with brief excursions to ~0.7 N during aggressive cornering.
- Collision: Maintained inter-agent clearance above 0.67 m (limit 0.8 m) for $N=3$ .
- Swing Rate: Strictly held within the 1.5 rad/s limit.
- Tilt: Operated near the 28.6° limit, slightly exceeding it under load (max 32°).
- Cable Angle: Exceeded the 34.4° limit (reaching 48°) during aggressive cornering, identified as a relative-degree-two output where the force-level filter has limited authority.
Ablation Studies: Removing the ESO feedforward doubled the tracking error (+110%) and allowed cable angles to reach 68°. Removing the CBF filter increased error by 19% and reduced clearance to 0.32 m. The concurrent learning feedforward had a neutral effect on tracking RMSE (as the integrator absorbs static deficits) but was essential for implicit coordination.
Scalability: The per-agent controller remained structurally reusable for $N \in \{2, \dots, 6\}$ . However, safety margins (clearance) degraded significantly for $N \ge 4$ due to fixed formation geometry, not control law limitations.

Significance and Claims
The paper claims to close the gap between geometric manifold control, decentralized adaptive estimation, and runtime safety supervision. It demonstrates that a team of UAVs can achieve coordinated transport without a central coordinator, without knowing the team size or payload mass, and without exchanging complex state information. The architecture successfully handles flexible cables and wind disturbances in a sensor-in-the-loop simulation.

The authors modestly note limitations:

The decentralized single-cable load-position estimate is poor due to observability issues (though it does not affect the tracking loop).
The sequential CBF projection guarantees priority-ordered feasibility rather than full ISSf when multiple constraints activate simultaneously.
The cable angle constraint is not strictly enforced during aggressive maneuvers.
The results are currently simulation-only; hardware flight experiments are identified as the necessary next step.

The work establishes that geometric decomposition, combined with local adaptive estimation and layered safety supervision, provides a robust framework for safety-critical cooperative aerial transport.