In modern naval warfare, fixed-wing drones play a pivotal role in reconnaissance, strike assessment, and ballistic correction. Their ability to operate from aircraft carriers extends their operational range far beyond their endurance. However, landing a fixed-wing drone on a moving deck presents unique challenges: the landing platform moves rapidly, the drone must maintain a minimum airspeed to generate lift, and it cannot brake effectively in mid-air. To address these difficulties, we propose a segmented action guidance strategy that divides the autonomous landing process into two main phases: transition guidance and approach guidance. The transition phase uses a finite state machine to bring the drone from any initial state to a specific distance behind the moving deck. The approach phase employs action sampling to generate and evaluate motion primitives in real time, guiding the drone precisely to the touchdown point. Numerical simulations and hardware-in-the-loop experiments demonstrate that this strategy can land a fixed-wing drone on a carrier deck moving at various speeds, even with deck heave, and achieve landing errors within 2 m.
The fundamental challenge for fixed-wing drone landing on a mobile platform lies in the drone’s inability to hover or decelerate abruptly. Unlike rotary‑wing drones, a fixed-wing drone relies on forward speed to produce lift, and its descent path is constrained by the need to avoid stall and excessive sink rates. Moreover, the landing platform is a cooperative but moving target; the drone must continuously predict its future position and adjust its trajectory accordingly. Traditional trajectory tracking methods, which rely on pre‑planned reference paths, are insufficient in dynamic environments. Therefore, we adopt a reactive planning approach that periodically samples feasible actions and selects the one that best aligns with the landing goal.
To formulate the problem, we define three coordinate frames: the inertial frame Fe, the carrier frame Fc, and the drone body frame Fb. The drone’s state is represented in the carrier frame to simplify the landing objective. The Dubins airplane model is used for the fixed-wing drone’s kinematics, which assumes coordinated turns and a constant airspeed assumption. The model is:
$$
\begin{cases}
\dot{x} = V \cos\psi \cos\gamma \\
\dot{y} = V \sin\psi \cos\gamma \\
\dot{z} = -V \sin\gamma \\
\dot{\psi} = \omega_{\psi} \\
\dot{\gamma} = \omega_{\gamma}
\end{cases}
$$
where $V \in [V_{\min}, V_{\max}]$ is the ground speed, $\gamma$ is the flight path angle, $\psi$ is the heading angle, and $\omega_{\psi},\omega_{\gamma}$ are bounded angular rates. The turn is coordinated via roll, with relationship $\dot{\psi} = \frac{g}{V} \tan\phi$, where $\phi$ is the roll angle. The transition from the inertial frame to the carrier frame involves a rotation by the carrier’s yaw angle and a translation by the carrier’s position. This transformation allows us to treat the landing as a problem of guiding the fixed-wing drone to the origin of the carrier frame with zero lateral offset and heading aligned to the carrier’s forward direction.
We divide the landing process into two distinct stages. The first stage, transition guidance, is responsible for reducing altitude and positioning the fixed-wing drone roughly behind the carrier deck. The second stage, approach guidance, performs precise adjustments using an action‑sampling scheme.
Transition Guidance Stage
When the fixed-wing drone receives the landing command, it may be at an arbitrary altitude and position relative to the carrier. A finite state machine (FSM) determines the appropriate action based on the drone’s state. The states include:
- Loiter: maintain altitude when height is too high or when waiting for command.
- Descend: reduce altitude until a safe ceiling is reached.
- Fly to Rear: navigate to a point behind the carrier deck, typically at a distance that ensures sufficient glide distance.
- Turn Around: if the drone’s heading is opposite to the carrier’s direction, execute a turn until the heading difference is less than $\pi/2$ rad.
- Go Around: if the approach is predicted to fail, climb to a safe altitude and re‑enter loiter.
The FSM outputs constant commands for desired heading rate $\omega_{\text{cmd}}$, speed $V_{\text{cmd}}$, and flight path angle $\gamma_{\text{cmd}}$. For example, in the “Fly to Rear” state, $\psi_{\text{cmd}} = -\pi$ (in carrier frame), $\gamma_{\text{cmd}} = 0$. This stage is designed for robustness and simplicity, and it brings the fixed-wing drone to a region where precise guidance can take over.
Approach Guidance Stage
Once the fixed-wing drone is approximately aligned behind the carrier (heading difference < $\pi/2$ and adequate distance), the approach algorithm is activated. This algorithm is an extension of our earlier work; it incorporates heading alignment cost and operates directly in the carrier frame. At each planning cycle, we compute the dynamic window $V_r = V_s \cap V_d \cap V_a$, where:
- $V_s$: feasible speed and heading rate from drone performance constraints.
- $V_d$: reachable within one time step given acceleration limits.
- $V_a$: admissible region that ensures the heading points toward the carrier (avoid circling).
Sampling within $V_r$, we generate multiple motion primitives by forward‑simulating the Dubins model for a prediction horizon $\Delta T$. Each primitive is a sequence of states. The endpoint of each primitive is evaluated using a cost function:
$$
J = \lambda_y J_y + \lambda_{\psi} J_{\psi} + \lambda_V J_V + \lambda_t J_t
$$
where:
- $J_y = |y|$: lateral offset in carrier frame.
- $J_{\psi} = |\psi|$: heading misalignment.
- $J_V = |V – V_{\text{ref}}|$: speed deviation from desired approach speed.
- $J_t = |t_H – t_i|$: time heuristic difference, where $t_H$ is estimated time to reach the deck based on Dubins distance and speed, and $t_i$ is the ideal time to descend from current height at desired sink rate.
The heuristic distance $d_H$ is computed using Dubins path geometry, which accounts for the minimum turn radius $R$. For instance, when the fixed-wing drone is far laterally ($|y_{c1}| > R$), the path consists of an arc, a straight line, another arc, and a straight line; the distance is given by:
$$
d_H = (|\psi| + \pi)R + |y_{c1}| – 3R – x_{c1}
$$
When the drone is closer, other formulas apply. The ideal time $t_i = h / (V \sin\gamma_i)$ with $\gamma_i$ the optimal glide slope. The planner dynamically adjusts the desired flight path angle $\gamma_{\text{cmd}}$ based on the comparison between available distance and required distance. If distance is excessive, the drone reduces descent rate; if insufficient, it increases descent rate or performs a detour to lose altitude. A hysteresis threshold $t_{r1}$ and $t_{r2}$ prevents oscillation between modes.
Numerical Simulation Results
We conducted extensive numerical simulations using a Cessna‑like model with parameters listed in Table 1. The carrier moved at speeds between 3 m/s and 15 m/s, and deck heave (amplitude 1.52 m, frequency 0.6 rad/s) was added in the last 12 seconds. Five sets of tests were executed:
- Varying lateral offset ($-100$ to $100$ m) at constant heading and height.
- Varying initial heading ($-\pi/2$ to $\pi/2$).
- Varying initial height (50–100 m).
- Varying carrier speed (3–15 m/s).
- Carrier heave with compensation.
Table 2 summarizes typical touchdown errors. In all cases, the fixed-wing drone landed within 2 m of the desired touchdown point, well within the typical requirement of 12.2 m. The approach algorithm successfully handled initial misalignments, excessive altitude, and moving platforms.
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| $V_{\min}$ | 27 m/s | Max glide angle $\gamma_i$ | -0.07 rad |
| $V_{\max}$ | 40 m/s | Turn radius $R$ | 160 m |
| $a_{\min}, a_{\max}$ | ±1 m/s² | Time step $\Delta t$ | 0.05 s |
| $\omega_{\min}, \omega_{\max}$ | ±0.2 rad/s | Prediction horizon $\Delta T$ | 2 s |
| $\alpha_{\psi\min}, \alpha_{\psi\max}$ | ±0.2 rad/s² | Weights $\lambda_y,\lambda_{\psi},\lambda_V,\lambda_t$ | 1,25,0.5,10 |
| Speed resolution | 0.02 m/s | $t_{r1}, t_{r2}$ | 1 s, -0.5 s |
| Angular resolution | 0.002 rad/s |
| Lateral offset (m) | UAV position (x,y,z) (m) | Carrier x (m) |
|---|---|---|
| -100 | (187.78, 0.09, 0.10) | 189.25 |
| -60 | (185.93, 0.09, 0.11) | 187.50 |
| -20 | (188.48, -0.09, 0.11) | 190.00 |
| 20 | (188.48, 0.09, 0.11) | 190.00 |
| 60 | (185.93, -0.09, 0.11) | 187.50 |
| 100 | (187.78, -0.09, 0.10) | 189.25 |
When carrier heave was introduced, the planner predicted the deck’s vertical motion and adjusted the target height accordingly. The resulting touchdown errors remained below 2 m in all cases. For example, at carrier speed 3 m/s, the drone landed at (140.70, -0.05, -1.44) while the carrier’s vertical position was -1.49 m, giving an vertical error of 0.05 m.
Hardware-in-the-Loop Simulation
To validate the algorithm under realistic latency and communication constraints, we conducted hardware‑in‑the‑loop (HIL) simulations using a Pixhawk flight controller, FlightGear software, and an on‑board computer running the guidance algorithm under ROS. The full landing sequence was executed: initial loiter at 1000 ft, descent to 100 m, transition to rear of carrier, turn‑around, and final approach. The carrier moved at 20 kn (10.28 m/s) with a heading of 200°. The entire process from command to touchdown took about 204 s. At touchdown, the fixed-wing drone’s relative position in the carrier frame was (-0.224, -1.985, 0.086) m, with a heading error of 0.3°. The results confirm that the segmented strategy works in a realistic simulation environment with real control loops.
Conclusion
We have presented a segmented action guidance strategy for autonomous shipborne landing of fixed-wing drones. By decomposing the landing into transition and approach phases, and employing a finite state machine for rough guidance and action sampling for precise control, the system can handle a wide range of initial conditions and moving platform dynamics. Numerical simulations and hardware‑in‑the‑loop tests demonstrate that the algorithm achieves landing errors within 2 m on a moving carrier deck, even with heave motion. The approach is also applicable to net‑recovery and line‑recovery scenarios for fixed-wing drones, as the principles are similar. Future work will focus on extending the method to multiple‑drone coordination and handling more severe sea states.