⚡ electrical engineering

Feasibility and Singularity in High-Order Safety-Critical Control for Quadrotor UAVs

This paper proposes a robust, torque-aware high-order control framework for quadrotor teams that utilizes Gaussian processes to learn dynamic residuals and employs a fourth-order barrier function to prevent singularity and ensure collision avoidance under bounded inputs.

Original authors: Omayra Yago Nieto, Leonardo Colombo

Published 2026-09-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Omayra Yago Nieto, 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 swarm of small flying robots, each no larger than a dinner plate, tasked with navigating a crowded space together. These machines, known as quadrotors, are the workhorses of modern aerial robotics, used for everything from inspecting bridges to delivering medical supplies. However, they possess a fundamental physical limitation: they can only push themselves forward or backward by tilting their bodies and spinning their propellers to generate thrust in a single direction relative to their own frame. This makes them inherently difficult to control when they need to move sideways or stop suddenly without tilting. When a team of these robots flies in close proximity, the risk of collision is constant. To keep them safe, engineers use mathematical safety filters that act like an invisible guardrail, constantly checking the robots' paths and intervening only when a crash is imminent. The goal is to let the robots follow their intended course as closely as possible, stepping in only when absolutely necessary to prevent disaster.

The challenge intensifies when these robots must operate under strict physical limits, such as a maximum amount of force their motors can produce. In complex scenarios, the geometry of the situation can conspire against the safety system. If two robots approach each other from a specific angle, the direction they need to push to avoid a collision might be perpendicular to the direction their motors can actually push. In such moments, the safety system loses its ability to generate the necessary force, a phenomenon known as a singularity. Furthermore, even if the safety system works perfectly for each pair of robots individually, the combined demands of avoiding collisions with multiple neighbors at once might exceed the total force the entire team can generate. This creates a situation where the safety rules are mathematically impossible to satisfy simultaneously, leading to a breakdown in the system's ability to keep everyone safe.

In a recent study, researchers Omayra Yago Nieto and Leonardo Colombo investigated these specific failure modes in teams of quadrotor drones. They sought to understand exactly when and why these safety filters fail and to design a new control method that could overcome these limitations. Their work focuses on two distinct problems: the loss of effectiveness when a robot's available thrust is misaligned with the needed direction, and the collective inability of a group to satisfy all safety constraints at once due to shared power limits. By analyzing the physics of the situation, they identified that the standard safety checks, which look at the distance between robots, often fail because they do not account for the robot's ability to rotate its body to change the direction of its thrust.

To solve this, the researchers developed a more sophisticated approach that looks further ahead in time. Instead of just checking the current distance between robots, their new method considers how the robots' speed, acceleration, and rotational forces will evolve over the next few moments. They introduced a technique that explicitly includes the torque—the twisting force used to rotate the drone—into the safety calculations. By doing so, they created a safety filter that understands that even if a robot cannot push directly away from a collision right now, it can rotate its body to point its thrust in the right direction a fraction of a second later. This "torque-aware" extension ensures that the safety system always has a valid way to generate a corrective force, provided the drone is still producing some thrust.

The researchers also tackled the problem of collective feasibility. They created a mathematical measure to determine if the entire team can satisfy all their safety constraints at the same time. Their analysis showed that even when individual pairs of robots are not in a singular, unmanageable state, the group as a whole might still be in a situation where no single set of commands can keep everyone safe. To address this, they formulated a new control algorithm that constantly checks this group-wide feasibility. If the system detects that the team is approaching a point where safety becomes impossible, the algorithm adjusts the robots' behavior to stay within a safe operating zone, ensuring that a valid solution always exists.

To handle real-world unpredictability, such as wind gusts or slight variations in the robots' weight, the team incorporated a learning component. They used a statistical method called Gaussian processes to learn the unknown errors in the system's model directly from data. Rather than trying to calculate every possible disturbance, the system learns the pattern of these errors and builds a safety margin around them. This allows the controller to remain robust even when the environment behaves in ways the original model did not predict. The result is a safety filter that is both precise and adaptable, capable of maintaining safe distances even in the face of uncertainty.

The team tested their new controller in a computer simulation involving three quadrotors performing a complex crossing maneuver. In this scenario, the robots had to fly past each other in a tight formation, a situation designed to trigger the very failures the researchers aimed to prevent. When they ran the simulation with a standard, older safety controller, the system failed. The robots came dangerously close, with the distance between them dropping to just 0.300 meters, far below the safe limit of 0.8 meters. The older controller became mathematically stuck, unable to find a solution that would keep the robots apart, and the simulation had to be stopped.

In contrast, when the researchers applied their new torque-aware controller, the robots completed the maneuver flawlessly. The minimum distance between any two drones during the entire flight was 0.822 meters, safely above the required threshold. The new system successfully navigated the robots through the tightest part of the crossing without ever losing control or violating the safety constraints. The simulation also revealed that while the older controller failed because it could not account for the robots' ability to rotate, the new controller used this rotational capability to generate the necessary avoidance forces. Even when the researchers introduced random disturbances to mimic wind and mechanical imperfections, the new system maintained a safe distance of 0.895 meters, proving its resilience.

The study concludes that by looking deeper into the physics of the drones—specifically by including their rotational dynamics and learning from data—engineers can create safety systems that are far more reliable. The researchers demonstrated that the fear of safety filters failing due to geometric singularities or collective power limits can be mitigated by a control strategy that respects the full capabilities of the machine. Their work provides a concrete path forward for deploying teams of drones in crowded, real-world environments where safety is non-negotiable, ensuring that these aerial robots can work together without crashing, even when the situation seems impossible.

Technical Summary: Feasibility and Singularity in High-Order Safety-Critical Control for Quadrotor UAVs

Problem Statement
This paper addresses the safety-critical control of teams of quadrotor UAVs operating under bounded inputs and pairwise collision-avoidance constraints. Quadrotors are intrinsically underactuated, generating translational force only along their body-fixed vertical axis. Consequently, collision avoidance depends heavily on instantaneous attitude and relative geometry. The authors identify two distinct failure modes in existing safety filters:

Pairwise Thrust Singularity: In specific geometric configurations where the relative displacement between two UAVs is orthogonal to their available thrust directions, the thrust magnitude cannot instantaneously modify the safety constraint (specifically, the second-order derivative of the distance barrier). This leads to a loss of input effectiveness.
Aggregate Infeasibility: Even when individual pairwise constraints are nonsingular, the simultaneous satisfaction of all collision-avoidance constraints may be impossible due to shared bounded input limits. This "joint infeasibility" occurs when the required control effort to satisfy all constraints exceeds the available thrust bounds, even if no single constraint is singular.

Methodology
The proposed approach integrates high-order control barrier functions (HOCBFs) with a torque-aware dynamic extension and Gaussian Process (GP) learning.

Thrust-Based Analysis: The authors first characterize the thrust singular set and define a pairwise thrust-effectiveness measure. They introduce an exact scalar "aggregate feasibility margin" ( $\Gamma^{rob}_{th}$ ) to quantify the simultaneous feasibility of all pairwise inequalities under shared bounded thrust inputs.
Torque-Aware Dynamic Extension: To resolve the thrust singularity, the authors extend the system dynamics. By differentiating the pairwise distance barrier four times, they expose the attitude torques ( $\tau$ ) and thrust rates ( $\dot{f}$ ) as control inputs. A key theoretical result demonstrates that while thrust-acceleration coefficients may vanish at a thrust singularity, the torque coefficients remain non-zero (provided at least one UAV has non-zero thrust). This ensures the extended-input row does not vanish, restoring control authority over the safety constraint.
Robustness via Gaussian Processes: Instead of differentiating unknown perturbations (which would require higher-order derivatives of uncertain terms), the authors use GPs to directly learn the fourth-order HOCBF residual. This provides a uniform robust margin without requiring explicit differentiation of the uncertainty.
Control Formulation: The final controller is a Quadratic Program (QP) that minimizes the deviation from a nominal input while satisfying the robust fourth-order HOCBF constraints and actuator bounds. The QP includes conditions for both pairwise nonsingularity (via the dynamic extension) and local aggregate feasibility.

Key Contributions

Characterization of Failure Modes: The paper distinguishes between the loss of pairwise thrust effectiveness (a geometric singularity) and the joint incompatibility of nonsingular inequalities under bounded inputs (an aggregate feasibility issue).
Torque-Aware Dynamic Extension: The authors propose a fourth-order barrier construction that explicitly incorporates attitude torques. They prove that this extension prevents the input coefficient from vanishing under positive thrust bounds, effectively bypassing the thrust singularity.
Exact Aggregate Feasibility Margin: A scalar measure is derived to determine the feasibility of the entire set of collision constraints simultaneously, offering a rigorous condition for when a safe control input exists.
GP-Based Residual Learning: The method employs GPs to learn the fourth-order residual directly, providing robust safety margins without the need to differentiate unknown perturbations.

Results
Simulations were conducted on a three-quadrotor crossing maneuver designed to expose both thrust singularities and aggregate infeasibility.

Baseline Failure: A nominal geometric tracking controller violated the safety distance (reaching 0.300 m). A standard second-order thrust-only HOCBF-QP became infeasible at $t=4.53$ s, even when agents were still 1.673 m apart, due to aggregate infeasibility.
Proposed Controller Performance: The proposed torque-aware controller successfully completed the maneuver with a minimum pairwise distance of 0.822 m (above the 0.8 m limit).
Singularity Resolution: At a near-singular configuration, the thrust effectiveness measure dropped significantly ( $5.65 \times 10^{-2}$ ), but the torque-aware extended effectiveness remained high ( $2.57 \times 10^3$ ), confirming the dynamic extension's ability to maintain control authority.
Aggregate Feasibility: While the second-order approach showed a negative aggregate margin ( $\Gamma^{rob}_{th} = -2.815$ ), the proposed controller maintained a positive extended aggregate margin throughout the maneuver ( $\min \Gamma^{rob}_{ext} = 5278.6$ ).
Robustness: Under smooth, agent-dependent force and torque perturbations, the GP-based robust controller maintained the safety margin and kept all actuator commands within prescribed bounds.

Significance and Claims
The paper claims to provide a rigorous framework for safety-critical control in multi-agent quadrotor systems that accounts for both geometric singularities and the limitations of shared bounded inputs. By exposing attitude torques in a fourth-order barrier, the method ensures that safety constraints remain controllable even when thrust directions are geometrically unfavorable. The authors emphasize that pairwise nonsingularity alone is insufficient for safety; the aggregate feasibility of the entire team must be guaranteed. The work concludes that the resulting QP guarantees collision avoidance and recovers the nominal input whenever it satisfies the robust safety and actuator constraints, provided the system remains persistently feasible. Future work is noted to focus on experimental validation and less conservative feasibility guarantees.

Technical Summary: Feasibility and Singularity in High-Order Safety-Critical Control for Quadrotor UAVs

More like this