Embedded Safe Reactive Navigation for Multirotors Systems using Control Barrier Functions

TL;DR

This work addresses collision avoidance for multirotors in unknown environments by formulating obstacle avoidance as a composite Control Barrier Function (CBF) that relies solely on onboard range measurements. The authors integrate a safety filter into the PX4 open-source autopilot, computing a safe acceleration command via a QP that minimally deviates from the nominal reference while ensuring collision avoidance and limited field-of-view violations. They extend CBF theory to handle multiple constraints (through ECBFs and a soft-min composite) and provide a practical embedded implementation with real-time performance on a small UAV, including two hardware experiments demonstrating adversarial safety and robust trajectory tracking. The approach enables safe, mapless navigation with low computational burden and direct hardware integration, making it suitable for widespread adoption in autonomous aerial robotics.

Abstract

Aiming to promote the wide adoption of safety filters for autonomous aerial robots, this paper presents a safe control architecture designed for seamless integration into widely used open-source autopilots. Departing from methods that require consistent localization and mapping, we formalize the obstacle avoidance problem as a composite control barrier function constructed only from the online onboard range measurements. The proposed framework acts as a safety filter, modifying the acceleration references derived by the nominal position/velocity control loops, and is integrated into the PX4 autopilot stack. Experimental studies using a small multirotor aerial robot demonstrate the effectiveness and performance of the solution within dynamic maneuvering and unknown environments.

Figures (9)

• Figure 1: Graphical explanation of the composite CBF $h(\mathbf{x})$ constructed from $h_i(\mathbf{x})$ for $i \in \{1,3\}$. $h(\mathbf{x})$ is a smooth under-approximation of the function $\min_i \space h_i(\mathbf{x})$. For increasing the value of $\kappa$ (shown with decreasing opacity of $h$), the approximation becomes less conservative.
• Figure 2: Problem definition and frame conventions used in this work. The multirotor aerial robot seeks to avoid all visible obstacles present in the current sensor measurement. The inertial frame is denoted as $\mathcal{I}$, while the body frame is denoted as $\mathcal{B}$.
• Figure 3: Simplified planar navigation example with a constrained field of view. It illustrates that the vehicle (black) with instantaneous velocity $\mathbf{v}$ will approach the obstacles (red) located just outside the fov if the angles $\phi_1$, $\phi_2$ are less than $\pi/2$.
• Figure 4: Cascade control scheme diagram with the safety filter introduced.
• Figure 5: Safety filter response for different parameter values.