⚡ electrical engineering

Safety-Critical Control for Quadrotor UAVs via Decentralized Navigation Functions

This paper proposes a safety-critical control framework for teams of quadrotor UAVs that utilizes decentralized navigation functions to generate translational references and employs an aggregated robust HOCBF-QP filter to ensure high-probability collision avoidance despite model uncertainty and actuation constraints.

Original authors: Omayra Yago Nieto, Alexandre Anahory Simoes, Leonardo Colombo

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Omayra Yago Nieto, Alexandre Anahory Simoes, Leonardo Colombo

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 world where swarms of tiny, buzzing drones are the new delivery drivers, aerial photographers, and search-and-rescue heroes. To make this happen, we need them to work together like a perfectly choreographed dance troupe, never bumping into each other while zooming toward their goals. This is the realm of safety-critical control, a branch of robotics dedicated to ensuring machines don't crash, even when things go wrong.

The challenge is that drones are a bit clumsy in their thinking. They have a "brain" that can calculate the perfect path to avoid a tree or a friend, but their "muscles" (the motors) are limited. They can only push air straight down to move, meaning they can't just slide sideways like a car; they have to tilt their whole bodies to go where they want. Furthermore, the real world is messy. Wind gusts, heavy payloads, or just a slightly broken motor can make the drone's actual movement differ from what its brain predicted. This paper tackles the tricky problem of teaching a swarm of these underpowered, unpredictable drones how to follow a perfect, collision-free dance routine without actually crashing, even when the wind is blowing and their internal maps are slightly fuzzy.

The Dance of the Clumsy Drones

In this study, researchers Omayra Yago Nieto, Alexandre Anahory Simoes, and Leonardo Colombo tackled a specific headache: How do you tell a team of quadrotor drones (the kind with four spinning propellers) to move safely when their "ideal" instructions don't match their physical reality?

Think of the drones as a group of dancers who have been given a script. The script, generated by something called Decentralized Navigation Functions, tells each dancer exactly where to go to avoid bumping into others and to reach the stage center. It's like a GPS that knows exactly where everyone else is and says, "You, move left; you, move right."

But here's the catch: These drones are underactuated. In plain English, they are like a unicycle that can only pedal forward or backward; it can't just slide sideways. The script might say, "Move diagonally," but the drone can only generate thrust (push) straight down through its body. To move diagonally, it has to tilt its whole body. This tilt-and-push method introduces a gap between the "perfect script" and the "messy reality." If the drone tilts too slowly or the wind pushes it, it might miss the mark and crash into its dance partner.

The Safety Net: A "Smart Filter"

To fix this, the authors built a safety filter that acts like a vigilant referee standing between the dance script and the drone's motors. This referee uses a mathematical tool called a High-Order Control Barrier Function (HOCBF).

Imagine the HOCBF as a super-precise rulebook that checks the distance between every pair of drones. It doesn't just look at where they are now; it predicts where they will be in the next split second based on how fast they are moving and how hard they are pushing. If the rulebook sees a collision coming, it steps in.

However, the real world is unpredictable. The drones might face a sudden gust of wind or have a motor that isn't quite as strong as the manual says. To handle this, the team used Gaussian Processes (GPs), which are essentially a way for the drones to "learn" their own mistakes in real-time. As the drones fly, they collect data on how the wind is actually affecting them and update their internal models.

The paper's main innovation is combining these two ideas:

The Translation Error: They figured out exactly how much the "tilt-and-push" method messes up the perfect script. They calculated a specific "error bound," which is like a safety margin that says, "We know the drone might be off by this much because of how it tilts."
The Learning Safety Net: They plugged the "learning" data (the GP predictions) directly into the safety rulebook. Instead of just saying "Don't get closer than 1 meter," the rulebook now says, "Don't get closer than 1 meter plus whatever the wind might push you."

The "Aggregated" Solution

The team created a system called an Aggregated Robust HOCBF-QP Safety Filter. This is a mouthful, but think of it as a group decision-maker. Instead of each drone trying to solve the puzzle alone, they all send their current speed and position to a central calculator (the QP solver). This calculator looks at the entire swarm, checks the "learning" data for wind errors, and then makes the smallest possible tweak to the thrust (the push) of each drone to keep everyone safe.

It's like a conductor adjusting the volume of every instrument in an orchestra just enough to keep the music harmonious, without changing the melody. The goal is to change the drones' path as little as possible while guaranteeing they don't crash.

What the Simulations Showed

The authors tested their idea in a computer simulation with six drones performing a "viewpoint exchange" maneuver. This is like a camera crew where the drones fly in circles around a moving object, swapping places to get different angles. Their paths were designed to cross, which is a recipe for disaster if the safety system fails.

The results were promising:

The Baseline: When they used a standard controller that tried to compensate for wind but didn't have the special safety filter, the drones got too close. The minimum distance between them dropped to 1.1093 meters, which was less than the required 1.16 meters. They almost crashed.
The New System: When they turned on the new "Aggregated Robust HOCBF-QP" filter, the drones stayed safe. The minimum distance was maintained at 1.2540 meters.
The Cost: The filter had to tweak the drones' thrust (the power of their engines) to keep them safe. The average change in thrust was about 2.39 Newtons, with a maximum spike of 31.03 Newtons. This shows the system is effective but does require the drones to work a bit harder to stay safe.

They also compared their "robust" system (which accounts for learning errors) against a "non-robust" version (which ignores the learning uncertainty). The non-robust version kept the drones apart physically but violated the mathematical safety certificate, meaning it was "lucky" rather than "safe." The robust version, however, kept the mathematical safety certificate intact, proving that the system is reliable even when the wind is unpredictable.

The Bottom Line

This paper doesn't claim to have solved every problem in drone safety. The authors admit that their system relies on the math working out (the "QP" being feasible) and that there are tricky situations, like when two drones are perfectly perpendicular to each other's thrust, where the math gets harder. They also note that this was tested in simulation, not on real physical drones flying in a real park yet.

However, the study successfully bridges a major gap. It shows how to take a beautiful, perfect mathematical plan for a swarm of drones and make it work on real, clumsy, wind-blown machines. By combining a "learning" system that predicts errors with a "safety filter" that corrects them, they created a framework that guarantees, with high probability, that these aerial dancers will never bump into each other, even when the music gets chaotic.

Technical Summary: Safety-Critical Control for Quadrotor UAVs via Decentralized Navigation Functions

Problem Statement
The paper addresses the challenge of coordinating teams of quadrotor Unmanned Aerial Vehicles (UAVs) in safety-critical environments. While decentralized navigation functions (DNFs) provide a systematic method for generating goal-directed, collision-aware motion fields for multi-agent systems, a fundamental mismatch exists between these theoretical constructs and quadrotor dynamics. DNFs typically generate fully actuated translational reference forces, whereas quadrotors are underactuated, capable of producing thrust only along their body-fixed vertical axes. Furthermore, real-world deployment requires accounting for model uncertainties, such as unmodeled aerodynamic effects and external disturbances. The core problem is how to implement the ideal reference forces from DNFs into actual thrust and attitude commands for underactuated quadrotors while guaranteeing collision avoidance in the presence of learned model uncertainty.

Methodology
The proposed framework integrates three main components:

Thrust-Attitude Implementation of Navigation Forces:
The authors construct a control strategy to convert the fully actuated reference forces generated by DNFs into feasible thrust magnitudes and attitude commands. They define a desired total acceleration based on the negative gradient of the navigation function and learned model predictions. The nominal thrust is calculated as the orthogonal projection of the required force onto the current thrust direction ( $R_i e_3$ ). The paper derives an explicit error bound quantifying the discrepancy between the ideal fully actuated dynamics and the actual underactuated implementation, separating errors caused by underactuation (attitude tracking) from those caused by model estimation.
Learning-Based Uncertainty Modeling:
Unknown force and torque dynamics are estimated online using Gaussian Processes (GPs). The learning framework provides mean predictions of the disturbances and, crucially, high-probability confidence bounds on the prediction errors. These bounds are used to characterize the uncertainty in the system dynamics.
Aggregated Robust HOCBF-QP Safety Filter:
To enforce safety, the authors employ High-Order Control Barrier Functions (HOCBFs). Since collision avoidance constraints for quadrotors have a relative degree of two with respect to thrust inputs, standard Control Barrier Functions (CBFs) are insufficient. The authors formulate a pairwise HOCBF constraint that incorporates the learned GP uncertainty.
- A robustness margin is derived to account for the difference between the true system dynamics and the learned model.
- An aggregated Quadratic Program (QP) is solved at each time step to minimally modify the nominal thrust magnitudes. This optimization ensures that the robust HOCBF conditions are satisfied for all agent pairs, guaranteeing forward invariance of the safe set with high probability. The QP couples agents only through pairwise constraints, preserving the decentralized nature of the navigation planning while centralizing the safety enforcement.

Key Contributions
The paper makes three primary contributions:

Implementation and Error Quantification: It constructs a thrust-attitude implementation for decentralized navigation forces on quadrotor teams and derives an explicit error bound relative to the fully actuated reference dynamics.
Integration of Learning Uncertainty: It incorporates Gaussian Process force uncertainty directly into the pairwise relative-degree-two HOCBF constraints through an explicit robustness bound, rather than treating uncertainty as an additive disturbance without probabilistic guarantees.
Robust Safety Filter: It formulates an aggregated robust HOCBF-QP safety filter that minimally modifies nominal thrusts. This filter guarantees pairwise collision avoidance with high probability, provided the QP remains feasible.

Results
The methodology was validated through simulations involving a six-UAV team performing a cinematography-inspired reconfiguration task with crossing trajectories and unknown spatially varying wind disturbances.

Implementation Accuracy: The thrust-attitude implementation successfully tracked the reference forces, with the total implementation error remaining below the theoretical bound throughout the maneuver.
Safety Performance: The nominal controller (using GP compensation but no safety filter) violated the prescribed minimum separation distance ( $d_{min}$ ). In contrast, the robust HOCBF-QP maintained a minimum distance strictly greater than $d_{min}$ .
Robustness Verification: A comparison between a robust QP and a non-robust version (where the uncertainty margin was set to zero) demonstrated that the non-robust controller violated the tightened safety certificate (the GP-aware HOCBF condition), even though it appeared to maintain physical separation. The robust QP maintained the certificate deficit near zero, confirming the necessity of the uncertainty margin for probabilistic safety guarantees.
Feasibility: The QP remained feasible throughout the simulation with a 100% success rate, and the realized GP errors stayed within the adopted confidence bounds.

Significance and Claims
The paper claims to bridge a critical gap in safety-critical control for multi-quadrotor systems. While previous work has addressed decentralized navigation, underactuation, or learning-based control separately, this work unifies them. The authors emphasize that their approach allows for the direct application of fully actuated navigation forces to underactuated platforms while rigorously accounting for the implementation error and model uncertainty.

The significance lies in the ability to guarantee probabilistic collision avoidance without requiring conservative, worst-case bounds on uncertainty, which often lead to overly restrictive behavior. By using learned uncertainty bounds within a high-order barrier function framework, the system achieves a balance between aggressive goal-directed motion and rigorous safety. The authors note that the safety guarantee is conditional on the feasibility of the QP, acknowledging that singular configurations (where the relative position vector is orthogonal to thrust directions) may limit instantaneous control authority, a topic reserved for future work.