Landing Framework in UAVs & Air Traffic

Updated 7 July 2026

Landing frameworks are integrated architectures that couple sensing, state representation, decision-making, planning, and control to ensure safe touchdown in diverse domains.
They employ staged decomposition and decision–execution separation to balance precise perception with adaptive control, enhancing both site selection and trajectory planning.
Evaluations demonstrate high success rates and robustness through techniques like Bayesian inference, adaptive Kalman filtering, and cooperative multi-agent strategies under challenging conditions.

A landing framework is an integrated architecture that organizes sensing, state representation, decision-making, planning, and control around a landing task. In recent literature, the term is used across several problem classes: soft landing of resource-constrained UAVs on vertically oscillating platforms using optic flow and adaptive Koopman–MPC (Banday et al., 28 Jan 2025); evidence-based landing site selection and visual servoing in unstructured environments (Sajjadi et al., 2 May 2026); multimodal, neuro-symbolic, or monocular-3D landing-site assessment (Qian et al., 25 Oct 2025, Tan et al., 17 Jun 2025); fixed-time and cooperative landing on moving platforms (Izadi et al., 1 Jun 2026, Zhao et al., 6 Jan 2026); and, in air traffic management, probabilistic or scheduling-oriented frameworks for landing-time prediction and aircraft landing allocation (Kim et al., 9 Dec 2025, Maru, 18 Feb 2025). Across these usages, a landing framework is typically not a single controller but a multi-layer system that couples perception or estimation with downstream execution.

1. Meanings and scope of the term

In robotics and aerial autonomy, “landing framework” most often denotes a closed-loop pipeline for selecting, validating, and reaching a touchdown region. In air traffic management, the same term can denote a framework for predicting or scheduling landing events rather than controlling touchdown itself. The literature therefore uses the term at multiple abstraction levels, but consistently for architectures that integrate several subsystems rather than isolated algorithms (Maru, 18 Feb 2025, Kim et al., 9 Dec 2025).

Domain	Representative formulation	Typical components
UAV touchdown on moving platforms	Optic flow-guided adaptive landing; fixed-time dynamic landing; cooperative landing with active platform tilt	Estimation, trajectory generation, synchronization, feedback control
UAV landing-site assessment	Bayesian belief maps, monocular 3D safe-landing-zone estimation, neuro-symbolic reasoning	Perception, safety inference, feasibility filtering, ranking
Maritime landing	Stabilized deck plus UAV landing policy	Platform compensation, motion prediction, trajectory planning, touchdown control
Air traffic landing management	Probabilistic landing-time prediction, graph-enhanced scheduling	Interaction modeling, uncertainty quantification, sequencing

A recurrent distinction is between deciding where and when landing is acceptable and deciding how the vehicle should execute the final approach. Several frameworks make this separation explicit. The evidence-based UAV system in unstructured environments separates landing-site selection from visual-servo execution (Sajjadi et al., 2 May 2026), while LLM-grounded and neuro-symbolic systems separate semantic reasoning from MPC or rule-based decision layers (Cai et al., 9 May 2025, Qian et al., 2 Jul 2026). A plausible implication is that “landing framework” has become a term for modularity as much as for landing itself.

2. Recurring architectural patterns

A dominant pattern is staged decomposition. One class of frameworks uses a perception or estimation layer, then a planner, then a controller. The fixed-time dynamic landing architecture combines an adaptive unscented Kalman filter, a minimum-jerk trajectory generator with prescribed touchdown time, and nonlinear MPC (Izadi et al., 1 Jun 2026). The spectral-temporal maritime framework couples a motion predictor, an HPO-RRT* SE(3) planner, and a learning-augmented predictive controller (Liu et al., 17 Feb 2026). These systems formalize landing as sequential estimation, prediction, and control.

A second pattern is decision–execution separation. In evidence-based site selection, a Bayesian belief map over candidate regions determines the landing site, after which ORB feature tracking and image-based visual servoing perform alignment and descent (Sajjadi et al., 2 May 2026). In LLM-Land, a vision-language encoder and lightweight LLM with retrieval-augmented generation output semantic flags and safety parameters, which then modify the safe flight corridor for a nonlinear MPC rather than directly commanding the UAV (Cai et al., 9 May 2025). Neuro-symbolic systems similarly isolate perception from symbolic safety logic, then rank regions of interest under mission-dependent cost functions (Qian et al., 25 Oct 2025).

A third pattern is cooperative or decoupled multi-agent landing. The maritime platform framework with a USV “Nest” and UAV divides the problem between a 3-RPU stabilized deck controlled by SAC and a UAV multimodal RL policy for final approach (Neves et al., 30 Jun 2026). By contrast, the bi-directional adaptive framework treats the moving platform as an active agent that tilts its surface to create an optimal terminal attitude for the approaching quadcopter, thereby parallelizing alignment and descent (Zhao et al., 6 Jan 2026). These two approaches differ in implementation, but both reject the assumption that the landing surface is passive.

A fourth pattern appears in non-robotic landing domains. The graph-enhanced DRL formulation of the Aircraft Landing Problem uses a graph state encoder, actor–critic policy, and safety-aware assignment layer for runway scheduling (Maru, 18 Feb 2025), while the probabilistic multi-agent landing-time predictor uses scene embedding, agent attention, and Gaussian parameter decoding to produce distributions rather than point estimates (Kim et al., 9 Dec 2025). Here the “landing framework” is organizational and predictive rather than kinodynamic.

3. Perception and state representation

Landing frameworks differ sharply in what they choose to represent. In optic-flow-guided landing on a vertically oscillating platform, the controlled output is the scalar optic flow

$x(t) \triangleq \frac{v(t)}{h(t)},$

with downward velocity $v(t)$ , height $h(t)$ , and optic flow $x(t)$ . The framework explicitly does not estimate full height and velocity from vision; instead it regulates optic flow directly using a data-driven Koopman model and event-based adaptive MPC (Banday et al., 28 Jan 2025). This is a compact task-aligned representation rather than a full state reconstruction.

Other frameworks represent safety probabilistically. In evidence-based landing-site selection, each candidate region has a latent binary safety variable $S_{t,i}$ , and the belief

$b_{t,i} = p(S_{t,i}=1 \mid \mathbf{y}_{1:t,i})$

is updated recursively from flatness, slope, and obstacle cues, producing a temporally consistent belief map (Sajjadi et al., 2 May 2026). This representation treats safety as an inferred variable rather than a deterministic threshold test.

Monocular-3D frameworks instead predict geometry explicitly. VisLanding produces metric depth, surface normals, and a binary safe–unsafe landing map from a single RGB image, then estimates landing-zone area by combining depth and normals (Tan et al., 17 Jun 2025). Neuro-symbolic systems lift semantic segmentation into a probabilistic semantic scene graph whose nodes encode regions and attributes such as is_large_area, is_flat_surface, and is_accessible, while edges encode spatial relations such as adjacent_to, contain, and near_to (Qian et al., 2 Jul 2026).

Large-model systems add semantic context beyond local geometry. LLM-Land converts images into captions, retrieves safety rules, and outputs structured semantic variables such as is_dynamic and z_min, which then alter the feasible corridor of the controller (Cai et al., 9 May 2025). A related remote-sensing framework combines RS imagery with point-of-interest data and multimodal LLMs to rank emergency landing sites using semantic risks that geometric sensors cannot encode, such as proximity to schools, gas stations, or roads (Hua et al., 1 Feb 2026). This suggests that state representation in landing frameworks is increasingly task-specific: optic flow for timing-to-contact, probabilistic safety for site persistence, or semantic scene graphs for auditable reasoning.

4. Planning and control formulations

Control layers range from classical servoing to MPC and hierarchical RL. In the optic-flow moving-platform framework, an offline Koopman operator model linearizes the input–output relation from acceleration to optic flow, an online adaptation scheme compensates unknown platform motion and ground effect, and event-driven MPC regulates the optic-flow reference while minimizing computational overhead (Banday et al., 28 Jan 2025). The paper provides convergence analysis guaranteeing global convergence of the tracking error to a uniform ultimate bound and Zeno-free behavior.

When landing-site selection is decoupled from execution, visual servoing is common. In the evidence-based system, ORB features inside the selected region are tracked, their centroid is aligned to the image center, and an image-based visual servo law generates bounded translational velocity commands using the interaction matrix and its pseudoinverse (Sajjadi et al., 2 May 2026). The decision layer is frozen once a region passes the belief threshold and feasibility constraint.

Semantic frameworks typically inject reasoning into optimization rather than replace optimization. LLM-Land defines an unsafe region $\mathcal{B}$ from semantic outputs and shrinks the original safe corridor $\mathcal{P}_c$ into a modified feasible set $\tilde{\mathcal{P}}_c$ such that

$\tilde{\mathcal{P}}_c \subseteq \mathcal{P}_c,\qquad \tilde{\mathcal{P}}_c \cap \mathcal{B} = \emptyset,$

after which a 3D nonlinear MPC tracks a trajectory within that corridor (Cai et al., 9 May 2025). The semantic layer is supervisory; the controller remains model-based.

Moving-platform landing has produced several specialized variants. The fixed-time framework plans a minimum-jerk trajectory to a prescribed touchdown time and tracks it with NMPC (Izadi et al., 1 Jun 2026). The bi-directional adaptive framework first optimizes an “optimal, stable terminal attitude” in a simplified longitudinal model, then generates a full 3D minimum-jerk trajectory, while a cooperative platform tilts to realize the desired terminal geometry and an $v(t)$ 0 controller tracks the maneuver (Zhao et al., 6 Jan 2026). Maritime frameworks may instead plan in SE(3) under non-convex constraints with HPO-RRT* and use a learning-augmented predictive controller for execution (Liu et al., 17 Feb 2026).

Hierarchical RL offers another reduction. WaveLander maps a compact 4D platform-relative observation to a scalar vertical velocity reference, while a conventional low-level controller performs lateral tracking and attitude stabilization (Li et al., 1 Jul 2026). The control problem is thereby reduced to timing-aware vertical decision-making rather than full end-to-end flight control.

5. Safety, uncertainty, and interpretability

A common misconception is that landing frameworks are primarily geometric and deterministic. Many recent systems are explicitly probabilistic or adaptive. The evidence-based site selector uses Bayesian filtering with persistence parameter $v(t)$ 1 and a hard geometric feasibility predicate based on maximum inscribed landing radius, with $v(t)$ 2 in experiments, so safety is the conjunction of posterior belief and physical footprint (Sajjadi et al., 2 May 2026). In fixed-time dynamic landing, an adaptive unscented Kalman filter updates process and measurement noise statistics online to reflect time-varying sensing quality, rather than relying on fixed covariances (Izadi et al., 1 Jun 2026).

Another misconception is that learning-based landing necessarily means an opaque end-to-end policy. Neuro-symbolic landing-site assessment systems explicitly oppose that design choice. They construct probabilistic semantic scene graphs, apply Scallop rules such as “avoid water and obstacles; prefer large, flat, accessible regions,” and produce calibrated safety scores with ranked regions of interest and human-readable justifications (Qian et al., 25 Oct 2025). NEUROSYMLAND reports that symbolic reasoning contributes only a small fraction of end-to-end latency, while the main computational cost lies in perception and PSSG construction (Qian et al., 2 Jul 2026). This directly addresses the claim that interpretability must be traded against edge feasibility.

A further misconception is that platform motion is always treated as exogenous disturbance. Cooperative landing frameworks reject this. The maritime multi-vehicle system physically stabilizes the deck with a 3-RPU mechanism before the UAV performs final approach (Neves et al., 30 Jun 2026), and the bi-directional adaptive framework turns deck attitude into an optimization variable (Zhao et al., 6 Jan 2026). By contrast, optic-flow and fixed-time frameworks keep the platform passive but compensate its motion via estimation, adaptation, or prediction (Banday et al., 28 Jan 2025, Izadi et al., 1 Jun 2026). The controversy is therefore not whether platform motion matters, but whether it should be absorbed by the UAV alone or redistributed across the larger system.

6. Evaluation, limitations, and directions of development

Empirical evaluation varies by domain but consistently measures more than touchdown success alone. In maritime cooperative landing, the decoupled USV–UAV system achieved a 100% landing success rate across 15 trials, with mean stabilization efficacy of 87.8% and the landing surface within 1 degree of horizontal for 96% of mission duration in rough conditions (Neves et al., 30 Jun 2026). In dual-expert monocular landing, the scale-adaptive perception module reduced mean touchdown error to 2.53 m with 1.03 m standard deviation, versus 5.53 m and 5.60 m for near-only and far-only baselines, and achieved 100% mission success across all altitudes in the reported evaluation (Tasnim et al., 16 Dec 2025). In the graph-enhanced aircraft-landing scheduler, the learned policy yielded a 99.95% reduction in computational time compared to mixed-integer programming and 38% higher runway throughput over FCFS, while producing solutions within 1 second (Maru, 18 Feb 2025).

Landing-site assessment papers emphasize robustness and edge deployment. NEUROSYMLAND achieved 61 successful assessments across 72 simulated landing scenarios, outperforming four baselines with 37–57 successes, and profiled bounded CPU, GPU, memory, and power on a Jetson Orin Nano 8 GB platform (Qian et al., 2 Jul 2026). VisLanding emphasizes cross-domain generalization: although in-domain segmentation metrics are lower than strong 2D baselines, cross-domain performance remains substantially stronger because depth and normal predictions regularize the safe-zone estimate (Tan et al., 17 Jun 2025). The remote-sensing and MLLM framework shows that adding POI context substantially improves ranking of emergency landing sites over image-only reasoning, which suggests that global semantic context can change what counts as a safe site (Hua et al., 1 Feb 2026).

Dynamic moving-platform systems emphasize timing and prediction accuracy. The fixed-time landing framework reports improved platform velocity prediction accuracy relative to EKF- and UKF-based methods and repeatable landings in hardware, with 10 dynamic landing trials yielding a mean touchdown error of 0.0787 m (Izadi et al., 1 Jun 2026). SpecFuse reports 3.2 cm prediction error, 4.46 cm landing deviation, 98.7% simulation success, 87.5% real-world success, and 82 ms latency on embedded hardware (Liu et al., 17 Feb 2026). WaveLander, by contrast, frames landing as a hierarchical RL problem and demonstrates robust landing performance under randomized platform motion, with success rates that remain substantially above constant-descent baselines across 30°, 40°, and 60° tilt disturbance settings (Li et al., 1 Jul 2026).

The literature also makes the present limitations explicit. Several frameworks remain simulation-only or require further outdoor validation (Neves et al., 30 Jun 2026, Tasnim et al., 16 Dec 2025). Others depend strongly on segmentation quality or known visual markers (Qian et al., 25 Oct 2025, Zhao et al., 6 Jan 2026). LLM-based systems still face latency, formatting, and hallucination issues, which is why grounding through retrieval or symbolic programs is repeatedly emphasized (Cai et al., 9 May 2025, Hua et al., 1 Feb 2026). A plausible implication is that future landing frameworks will continue to combine explicit structure—belief maps, physics-aware predictors, rule-based safety constraints, geometric controllers—with selectively learned modules for perception, adaptation, or timing policy, rather than converging to a single end-to-end paradigm.