CurviTrack: Curvilinear Trajectory Tracking for High-speed Chase of a USV

TL;DR

CurviTrack tackles the challenge of autonomously landing a UAV on a moving USV during high-speed, curvilinear maneuvers without relying on inter-vehicle communication. It introduces a drag-aware linear USV model integrated into a Kalman filter and an MPC-based UAV controller, enabling reliable prediction, tracking, and landing on a moving platform using only visual pose estimates. The approach delivers substantial reductions in prediction error and uncertainty, improved tracking accuracy, and higher landing success across diverse real-world scenarios, highlighting its practicality for decentralized, robust marine multi-robot operations. Overall, the work advances autonomous UAV-USV collaboration by enabling efficient, contact-rich operations under communication-denied conditions and dynamic sea states.

Abstract

Heterogeneous robot teams used in marine environments incur time-and-energy penalties when the marine vehicle has to halt the mission to allow the autonomous aerial vehicle to land for recharging. In this paper, we present a solution for this problem using a novel drag-aware model formulation which is coupled with MPC, and therefore, enables tracking and landing during high-speed curvilinear trajectories of an USV without any communication. Compared to the state-of-the-art, our approach yields 40% decrease in prediction errors, and provides a 3-fold increase in certainty of predictions. Consequently, this leads to a 30% improvement in tracking performance and 40% higher success in landing on a moving USV even during aggressive turns that are unfeasible for conventional marine missions. We test our approach in two different real-world scenarios with marine vessels of two different sizes and further solidify our results through statistical analysis in simulation to demonstrate the robustness of our method.

Figures (7)

• Figure 1: A collage of various moments from the real-world experiments.
• Figure 2: The entire uav control architecture; the mpc landing controller (red block) is integrated into the MRS systembaca2021mrs (grey blocks) and supplies the desired reference (velocity $\mathbf{\dot{r}}_d=\dot{x}\dot{y}\dot{z}^T$ and heading rate $\dot{\eta}_d$). In the MRS system, the first layer containing a Reference tracker processes the desired reference and gives a full-state reference $\bm{\chi}$ to the attitude controller. The feedback Position/Attitude controller produces the desired thrust and angular velocities ($T_d$, $\bm{\omega}_d$) for the Pixhawk flight controller (Attitude rate controller). The State estimator fuses data from Odometry & localisation methods to create an estimate of the uav translation and rotation ($\mathbf{x}$, $\mathbf{R}$). The Vision-based Detector obtains the visual data from the camera and sends the pose information b of the usv to the mpc. The individual states are sent to their respective prediction models, and using these predictions, the MPC generates the desired control reference according to the cost function.
• Figure 3: The red path represents the unchanged path taken by the USV when no drag force is acting on it; whereas the blue path shows how the difference between $V_x$ and $V_{x'}$ produces a $V_y$ vector and therefore, causes significant drag that leads to the turning of the USV.
• Figure 4: Comparison between predictions made by our proposed approach and the state-of-the-art approach by Novák et. al. NOVAK2025120606.
• Figure 5: A randomly picked example triangle trajectory for comparative analysis. Visible deviations arise from the search phase performed after aborted landings.