In recent years, the development of micro air vehicles (MAVs) has attracted significant attention due to their potential in both civilian and military applications where conventional aircraft, including fixed-wing drones, are unsuitable. Among the various types of MAVs, non-fixed-wing configurations—such as rotary-wing and flapping-wing designs—have emerged as promising platforms for operating in low-Reynolds-number environments, confined spaces, and complex terrains. Unlike fixed-wing drones, which require runways or relatively open spaces for takeoff and landing, non-fixed-wing MAVs offer greater maneuverability and the ability to hover, perch, or land on uneven surfaces. However, achieving reliable autonomous landing remains a critical technical challenge. This paper reviews the state-of-the-art autonomous landing methods for non-fixed-wing MAVs, covering control strategies, guidance algorithms, and mechanical landing structures. We present a systematic comparison of various approaches, including classic PID control, vision-based navigation, fuzzy logic, Kalman filtering, tau theory, and machine learning. Furthermore, we discuss the mechanical design of landing gears inspired by biological systems. The aim is to provide a comprehensive overview that highlights the progress made and the remaining challenges, particularly in comparison with the well-established landing techniques for fixed-wing drones.
1. Introduction
The autonomous landing capability of MAVs is essential for safe recovery, recharging, and mission continuity. Fixed-wing drones have mature solutions for landing, such as glide slope control and parachute deployment, but these methods require large areas or specific infrastructure. In contrast, non-fixed-wing MAVs—encompassing quadrotors, hexarotors, coaxial rotors, and flapping-wing vehicles—demand more sophisticated strategies due to their complex dynamics and the need to land on diverse surfaces, including moving platforms, vertical walls, and tree branches. The growing interest in biomimetic flight has spurred research into landing methods that mimic birds and insects. This review focuses on the autonomous landing of non-fixed-wing MAVs, with particular attention to the differences from fixed-wing drones. By analyzing over 40 recent studies, we categorize the methods into sensor-based control, model-based guidance, and learning-based approaches. Additionally, we survey the progress in landing gear design specialized for non-fixed-wing MAVs, which is a crucial aspect often overlooked in fixed-wing drone literature.
2. Autonomous Landing Methods for Rotary-Wing MAVs
Rotary-wing MAVs, such as quadrotors, have achieved a high level of autonomy in landing due to their well-understood dynamics and the availability of reliable sensors. We
$$ \dot{\mathbf{x}} = f(\mathbf{x}, \mathbf{u}) $$
where
$$ \mathbf{x} = [x, y, z, \phi, \theta, \psi, \dot{x}, \dot{y}, \dot{z}, p, q, r]^T $$
represents the state vector. The control input
$$ \mathbf{u} = [u_1, u_2, u_3, u_4]^T $$
corresponds to rotor speeds. Autonomous landing for these vehicles can be broadly classified into PID-based, vision-based, and fusion-based methods.
2.1 PID Control and Its Variants
The classic proportional-integral-derivative (PID) controller remains popular due to its simplicity. A cascaded two-level PID structure is often employed: an outer loop for position and velocity control, and an inner loop for attitude and angular rate control. The control law can be expressed as:
$$ \mathbf{u}(t) = K_p \mathbf{e}(t) + K_i \int_0^t \mathbf{e}(\tau) d\tau + K_d \frac{d\mathbf{e}(t)}{dt} $$
where
$$ \mathbf{e}(t) $$
is the error between the desired and actual states. However, pure PID algorithms lack real‑time sensing adaptability. To overcome this, researchers have integrated fuzzy logic to adjust PID gains dynamically, improving performance under disturbances. The fuzzy PID approach tunes the gains based on error and error rate using a rule base, as shown in the following table.
| Method | Year | Key Feature | Advantage | Limitation |
|---|---|---|---|---|
| Two-level PID (fixed-point) | 2015 | Static target, no external sensors | Fast response, high precision | Low autonomy, no real-time adaptation |
| Fuzzy PID + vision | 2020 | Vision-based pose estimation | High precision, no GPS required | Longer processing time |
| PID + Kalman filter (Apriltags) | 2017 | Landmark tracking with filtering | Fast tracking, decent accuracy | Relies on visual landmarks |
2.2 Vision-Based Landing
Vision sensors provide rich information for landing guidance. One common approach uses fiducial markers (e.g., Apriltags) combined with extended Kalman filters (EKF) to estimate relative pose. The filter prediction and update steps are:
$$ \hat{\mathbf{x}}_{k|k-1} = \mathbf{F}_k \hat{\mathbf{x}}_{k-1|k-1} + \mathbf{B}_k \mathbf{u}_k $$
$$ \mathbf{P}_{k|k-1} = \mathbf{F}_k \mathbf{P}_{k-1|k-1} \mathbf{F}_k^T + \mathbf{Q}_k $$
$$ \mathbf{K}_k = \mathbf{P}_{k|k-1} \mathbf{H}_k^T (\mathbf{H}_k \mathbf{P}_{k|k-1} \mathbf{H}_k^T + \mathbf{R}_k)^{-1} $$
$$ \hat{\mathbf{x}}_{k|k} = \hat{\mathbf{x}}_{k|k-1} + \mathbf{K}_k (\mathbf{z}_k – \mathbf{H}_k \hat{\mathbf{x}}_{k|k-1}) $$
where
$$ \mathbf{z}_k $$
is the measurement from the camera. This approach enables landing on moving platforms such as cars traveling at up to 50 km/h. Another vision-based method uses stereo cameras for depth estimation, while others employ Harris corner detection and contour detection to estimate the platform’s location.
2.3 Multi-Sensor Fusion and Robust Control
To enhance robustness, researchers have combined GPS, vision, and ultrasonic sensors. For example, a hybrid strategy divides the landing process into three phases: satellite positioning, visual positioning, and ultrasonic ranging. In turbulent wind conditions, sliding mode control integrated with vision and GPS has been demonstrated to ensure stable landing on dynamic platforms. The boundary layer sliding mode controller is described by:
$$ \mathbf{s} = \mathbf{e} + \lambda \int \mathbf{e} dt $$
$$ \mathbf{u} = -\mathbf{K} \cdot \text{sat}(\mathbf{s}/\Phi) $$
where
$$ \text{sat}(\cdot) $$
is the saturation function and
$$ \Phi $$
is the boundary layer thickness. This method shows strong robustness but requires accurate state estimation.
3. Autonomous Landing Methods for Flapping-Wing MAVs
Flapping-wing MAVs (FMAVs) are inspired by flying animals and offer unique advantages in agility and energy efficiency at low Reynolds numbers. However, their unsteady aerodynamics make autonomous landing far more challenging than for rotary-wing MAVs or fixed-wing drones. Unlike fixed-wing drones, FMAVs often need to perform vertical takeoff/landing or perch on surfaces. Research in this area is still at an early stage, with several emerging strategies.
3.1 Vertical Takeoff and Landing (VTOL) Capability
Some FMAV designs, such as those mimicking hummingbirds (e.g., Nano Hummingbird, Colibri) and dual‑wing configurations (e.g., NUS-Roboticbird), can hover and perform vertical landing. Their landing control can borrow PID and vision-based methods from rotary-wing MAVs, but the inherent instability of flapping flight requires higher control bandwidth. Table 2 summarizes the key FMAV platforms capable of VTOL.
| Platform | Year | Wing Type | Control Strategy | Landing Accuracy |
|---|---|---|---|---|
| Nano Hummingbird | 2012 | Single flapping (near horizontal) | PID-based | ~0.5 m |
| Colibri | 2017 | Twin flapping (horizontal) | Adaptive control | ~0.3 m |
| NUS-Roboticbird | 2018 | Dual flapping (horizontal) | PID with IMU | ~0.6 m |
3.2 Dynamics-Based Trajectory Planning
Chen et al. developed a short‑landing algorithm for bird‑scale FMAVs. They coupled Adams dynamics simulations with CFD to determine the relationship between flapping frequency and descent speed. The optimal flapping frequency
$$ f_{opt} $$
is found by minimizing the cost function:
$$ J = \int_0^{t_f} (v_x^2 + \alpha v_z^2 + \beta (f – f_0)^2) dt $$
subject to aerodynamic constraints. This model-based approach allows the FMAV to reduce its forward speed while maintaining descent stability. The landing algorithm is summarized in the flowchart presented in the literature (not shown here).
3.3 Machine Learning for Perching
Chirarattananon et al. employed adaptive tracking and iterative learning control (ILC) to achieve perching on vertical walls using a sub‑gram flapping-wing robot (RoboBee). The ILC update law is:
$$ \mathbf{u}_{k+1}(t) = \mathbf{u}_k(t) + \Gamma \mathbf{e}_k(t+ \Delta) $$
where
$$ k $$
is the iteration index,
$$ \Gamma $$
is the learning gain, and
$$ \Delta $$
accounts for the delay. After multiple trials, the robot learned a trajectory that allowed it to land on a wall with a success rate exceeding 90%. This method does not require an accurate dynamic model but demands repeated experiments.
3.4 Tau Theory-Based Guidance
Tau theory, originally proposed by Lee, provides a biologically plausible model for guidance. The tau function
$$ \tau(t) $$
is defined as the time‑to‑contact:
$$ \tau(t) = \frac{s(t)}{\dot{s}(t)} $$
where
$$ s(t) $$
is a gap variable (e.g., distance to the landing surface). Several guidance strategies based on tau have been developed for MAVs. The tau‑G strategy maintains a constant ratio between the tau of the gap and the tau of a guided variable, while tau‑H strategy couples harmonic motion. An improved tau‑H strategy was proposed to handle non‑zero initial and final velocities, which is essential for landing on moving platforms. The control law for tau‑H can be written as:
$$ \ddot{s} = \frac{1}{\tau_s} \left( \frac{\dot{s}^2}{s} – \frac{\dot{s}}{\tau_s} \right) $$
where
$$ \tau_s $$
is the tau of the gap. Kendoul developed the TauPilot autopilot system and successfully demonstrated vertical landing of a quadrotor using tau‑based guidance, showing its potential for future FMAV applications.
3.5 Motion Capture and Closed-Loop Control
Paranjape et al. used a Vicon motion capture system to provide real‑time state feedback for a tailless flapping‑wing vehicle. The vehicle’s wing dihedral angle was actively controlled to adjust flight path angle. The control law for the dihedral deflection
$$ \delta $$
is:
$$ \delta = k_\gamma (\gamma_d – \gamma) + k_\psi (\psi_d – \psi) $$
where
$$ \gamma $$
is the flight path angle and
$$ \psi $$
is the heading. This method achieved landing on a human hand with an error less than 0.5 m, demonstrating the feasibility of high‑precision landing using external tracking.
4. Landing Mechanical Structures
The landing gear design is critical for non‑fixed‑wing MAVs, especially for flapping‑wing platforms that lack the robust undercarriage of fixed‑wing drones or the simple skids of rotary MAVs. Inspired by birds and insects, researchers have developed several innovative mechanisms.
4.1 Jumping and Bouncing Mechanisms
Sivalingam designed a compression‑spring mechanism for flapping‑wing UAVs to enable jumping takeoff and bouncing landing. The leg stores energy during compression and releases it to propel the MAV upward. The dynamics are modeled as:
$$ m \ddot{h} + c \dot{h} + k h = F_{impact} $$
where
$$ m, c, k $$
represent mass, damping, and stiffness. Optimization using sequential quadratic programming yielded leg dimensions for a prototype. Later, Hudson et al. developed a four‑bar linkage bird‑foot mechanism that could both jump and absorb landing impacts.
4.2 Sliding Landing Gear
Inspired by fixed‑wing drones, some FMAVs use wheeled landing gear. Peterson et al. demonstrated successful sliding landing with a bird‑scale ornithopter. However, the periodic flapping motion induces vibrations that destabilize the wheels, requiring additional damping.
4.3 Biomimetic Claws for Perching
Roderick et al. developed SNAG (Stereotyped Nature‑inspired Adaptive Grasper), a robotic foot that can grasp branches upon impact. The claw converts kinetic energy into gripping force using a tendon‑based mechanism. The relationship between impact velocity
$$ v $$
and gripping force
$$ F_g $$
is:
$$ F_g = \frac{1}{2} \rho A C_d v^2 $$
where
$$ \rho $$
is air density,
$$ A $$
is effective area, and
$$ C_d $$
is drag coefficient. Gomez‑Tamm et al. used a single‑shape‑memory‑alloy (SMA) spring actuator to open and close a lightweight claw, enabling a 1.5 m wingspan FMAV to perch on various surfaces. These mechanisms bring non‑fixed‑wing MAVs closer to the versatility of fixed‑wing drones in complex environments.
5. Comparative Analysis and Discussion
To facilitate understanding, we summarize the key autonomous landing methods for non‑fixed‑wing MAVs in Table 3. The methods are compared in terms of underlying technology, advantages, and limitations. It is evident that while rotary‑wing MAVs have achieved practical autonomous landing, flapping‑wing MAVs still face significant hurdles in control stability and mechanical integration. Compared with fixed‑wing drones, non‑fixed‑wing MAVs offer superior spatial flexibility but require more sophisticated perception and actuation. Future research should focus on bridging the gap between rotary and flapping‑wing technologies, perhaps by adopting hybrid configurations. Additionally, the integration of neuromorphic vision sensors and deep reinforcement learning could lead to more robust and generalizable landing policies, finally enabling non‑fixed‑wing MAVs to match or exceed the autonomy of fixed‑wing drones in real‑world landing scenarios.
| Method | MAV Type | Core Technology | Advantages | Limitations |
|---|---|---|---|---|
| Two‑level PID | Rotary | PID control | Fast, simple | No sensing adaptation |
| Fuzzy PID + Vision | Rotary | Fuzzy logic, vision | No GPS, high precision | Longer computation |
| Apriltags + Kalman | Rotary | Visual landmarks, EKF | Fast tracking | Marker‑dependent |
| Binocular vision | Rotary | Stereo matching | Good depth accuracy | High computational cost |
| Sliding mode + visual | Rotary | SMC, GPS, vision | Robustness to wind | Requires precise state estimation |
| Ultrasonic localization | Rotary | One‑transmitter‑four‑receiver | Good accuracy in indoor | Limited range |
| Iterative learning | Flapping | ILC, adaptive control | Learns trajectories | Many trials needed |
| Tau‑based guidance | Rotary/Flapping | Tau theory | Bio‑inspired, no GPS | Lack of experimental validation for flapping |
| Motion capture + dihedral | Flapping | Vicon, dihedral control | High precision | Requires external infrastructure |
| Two‑stage trajectory optimization | Flapping | Data‑driven models | Short optimization time | Local minima issues |
6. Conclusion and Future Directions
Autonomous landing for non‑fixed‑wing MAVs has advanced considerably over the past decade. For rotary‑wing MAVs, mature solutions combining PID, vision, and Kalman filtering are already deployed in field applications. For flapping‑wing MAVs, progress in vertical takeoff/landing, tau guidance, and learning‑based perching has opened new possibilities. However, several challenges remain: (1) the need for lightweight and efficient landing mechanisms that can absorb impact without adding excessive mass; (2) the development of robust perception systems that work in GPS‑denied and low‑light environments; (3) the integration of flapping‑wing aerodynamics with control‑theoretic tools to achieve the same level of reliability as fixed‑wing drones. Future research should also explore multi‑modal sensor fusion and adaptive control strategies that can handle the unique dynamics of non‑fixed‑wing MAVs. By learning from the mature landing technologies of fixed‑wing drones, we can accelerate the transition of non‑fixed‑wing MAVs from laboratory prototypes to operational systems. The ultimate goal is to create MAVs that can autonomously land on any surface, under any weather condition, with the same ease as fixed‑wing drones on a runway.