Predictive MA-UAV Collaborative Control Framework

• The framework integrates multi-layer prediction with receding-horizon control and multi-agent reinforcement learning for real-time UAV coordination.
• It couples perception, localization, and secure communication modules to ensure uncertainty-aware control in dynamic environments.
• The approach supports adaptive mission planning, risk control, and reconfiguration, enabling scalable applications in smart agriculture and disaster response.

In the cited literature, predictive collaborative control for UAV systems is represented by architectures that integrate perception, prediction, decision-making, coordination, execution, and reconfiguration in UAV swarms, by receding-horizon control formulations, and by long-horizon sequential decision-making for uncertainty reduction (Jia et al., 2024, Zhang et al., 2023, Zhao et al., 4 Mar 2025). The record most directly labeled as a multi-UAV collaborative control framework, “EIA-SEC: Improved Actor-Critic Framework for Multi-UAV Collaborative Control in Smart Agriculture” (Zhou et al., 21 Dec 2025), does not actually present a usable smart-agriculture system model, reinforcement-learning formulation, EIA-SEC architecture, or experiments; accordingly, the topic is more accurately reconstructed from adjacent substantive works on cooperative swarm cognition, collaborative MPC, estimation-aware planning, secure communication, collaborative perception, and high-level multi-UAV task planning (Zhou et al., 21 Dec 2025).

1. Documentary Basis and Conceptual Scope

The most immediate clarification is documentary. The supplied record for (Zhou et al., 21 Dec 2025) is not a substantive technical manuscript: it is described as a generic LaTeX template with placeholder text, no smart-agriculture scenario description, no multi-agent formulation, no trajectory-planning algorithm, no EIA-SEC method, and no experiments. It therefore cannot serve as a primary technical source for a predictive MA-UAV collaborative control framework (Zhou et al., 21 Dec 2025).

The surrounding literature does, however, define a coherent field. One branch treats predictive collaboration as a cognitive swarm-management problem, exemplified by the Cooperative Cognitive Dynamic System (CCDS), which is organized around attention perception, learning and inference, and risk control, and implemented through infrastructure, abstraction, control, and application layers (Jia et al., 2024). Another branch formulates predictive collaboration directly as optimization or sequential decision-making: relative-motion nonlinear MPC in a non-inertial target frame (Zhang et al., 2023), nonlinear mixed-integer receding-horizon tracking of multiple maritime targets (Anastasiou et al., 25 Apr 2025), tightly joined factor-graph positioning-and-control (Yang et al., 2024), and COMA-based informative path planning under noisy inter-UAV communication (Zhao et al., 4 Mar 2025). A further branch emphasizes communication-driven anticipation, as in predictive movable-antenna positioning for secure UAV communications (Yu et al., 14 Aug 2025).

This suggests a broad but technically consistent interpretation of the topic. “Predictive” does not denote one single methodology. In the cited works it may mean explicit receding-horizon optimal control, long-horizon value optimization in MARL, future trend prediction in cognitive architectures, intent prediction for dynamic obstacles, or anticipatory positioning of communication hardware (Jia et al., 2024, Xu et al., 2024, Yu et al., 14 Aug 2025).

2. Architectural Decomposition of Collaborative Prediction and Control

A recurring architectural pattern is layered decomposition. CCDS is the clearest explicit statement of this pattern. At the cognitive level it contains three tightly coupled modules: attention perception, learning and inference, and risk control. At the implementation level it uses four layers: infrastructure, abstraction, control, and application. It further organizes swarm cognition through a perception-action cycle and a shunt cycle, the latter described as an internal mechanism for direct feedback among functional modules and compared to experience replay in reinforcement learning (Jia et al., 2024).

The operational significance of this decomposition is that collaborative control is not treated as trajectory generation alone. Attention perception performs semantic information measurement, demand-dependent perception, and multi-domain information analysis. Learning and inference includes continuous knowledge updating, adaptation from new information, prediction of future events or states, deduction, and internal self-inference. Risk control includes risk assessment, task switching, and strategy storage (Jia et al., 2024). In technical terms, the cited architecture embeds estimation, forecasting, supervisory switching, and reconfiguration within the control loop rather than treating them as peripheral utilities.

A complementary mission-layer decomposition appears in Agent4Drone. There the workflow is structured into observation, task understanding, memory, planning, execution, and verification, operating through bounded tool chains under partial observability and hidden validation logic (Zhang et al., 30 Jun 2026). This is not a low-level controller, and the paper explicitly states that it does not address flight stabilization, continuous trajectory optimization, or formal motion planning. Even so, its closed-loop structure mirrors a high-level receding-execution architecture: observe, decompose, assign, execute a bounded sequence, verify, and replan (Zhang et al., 30 Jun 2026).

A plausible implication is that a full predictive MA-UAV framework is naturally stratified. The cited literature supports at least four interacting layers: mission decomposition and assignment, perception and localization, predictive planning or control, and safety/reconfiguration supervision (Jia et al., 2024, Zhang et al., 30 Jun 2026).

3. Formal Predictive Formulations

The literature spans several distinct predictive formalisms.

Paradigm	Core formulation	Predictive quantity
COMA-based MARL IPP	CTDE MARL with global reward and counterfactual advantage	Future entropy reduction in belief maps
CoNi-MPC	Nonlinear MPC in the target’s non-inertial frame	Future relative pose and velocity
Maritime collaborative MPC	NMIP over UAV motion, FoV logic, and KF covariance propagation	Future estimation covariance
JPCM	Factor graph over states, controls, measurements, and MPC factors	Jointly inferred current and future states
Secure communication MPC	Finite-horizon trajectory plus beamforming and AN optimization	Future secrecy rate, path cost, and power
Movable-antenna predictive positioning	Supervised prediction of future optimal MA locations	Future communication-optimal array geometry

In COMA-based informative path planning, the global state is $s^t=\{\mathcal{M},p^t_{1:N},b\}$, the joint action is $u^t\in U^N$, and the reward is based on relative reduction in map entropy. The objective is to maximize cumulative entropy reduction over a finite budget, so prediction is expressed through long-horizon value estimation rather than through explicit MPC (Zhao et al., 4 Mar 2025). The paper is explicit that this is predictive in the sequential-decision sense, not in the classical receding-horizon sense.

CoNi-MPC is closer to classical predictive control. The UAV is controlled directly in the moving target frame $N$, with state built from relative position, relative velocity, relative orientation quaternion, target linear acceleration, target angular velocity, and target angular acceleration. Its nonlinear MPC minimizes state-reference deviation and input deviation from hover subject to the relative dynamics and actuator bounds (Zhang et al., 2023). The core relative-acceleration model explicitly includes Euler, Coriolis, and centrifugal terms: $${}^N\dot{\boldsymbol{v}}_B = -[{}^N\boldsymbol{\beta}_N]_{\times} {}^N\boldsymbol{p}_B -2[{}^N\widehat{\boldsymbol{\Omega}_N}]_{\times} {}^N\boldsymbol{v}_B -[{}^N\widehat{\boldsymbol{\Omega}_N}]_{\times}^2 {}^N\boldsymbol{p}_B + {}^N\boldsymbol{R}_B {}^B\boldsymbol{T}_B - {}^N\widehat{\boldsymbol{a}_N}.$$ This formulation removes dependence on world-frame state estimation for both agent and target, prior target motion models, and continuous global replanning for several relative-motion tasks (Zhang et al., 2023).

The maritime multi-target-tracking framework makes prediction estimator-centric. For planning agent $a$, the receding-horizon objective is

$$\mathcal{J}^a = \sum_{k=0}^{K-1}\sum_{j=1}^{\mathcal{C}^a}\operatorname{tr}\!\left(P^j_{\tau+k|\tau}\right),$$

so the controller optimizes future sensing geometry by minimizing predicted target covariance rather than direct geometric error (Anastasiou et al., 25 Apr 2025). The resulting NMIP couples UAV dynamics, field-of-view logic, pseudo-measurement generation, Kalman gains, posterior covariance updates, and collision constraints.

JPCM reformulates the estimation-control interface itself. Rather than fixing the current state from a prior positioning module and then running MPC, it constructs a single factor graph containing absolute and relative positioning factors, dynamic factors, reference-trajectory factors, and control-limit factors (Yang et al., 2024). In the reduced formulation, the joined objective is written as

$$f_{\mathrm{JPCM}}=f_{\mathrm{MPC}}+f_{\mathrm{dyn}}+\|H_0x_0-Z_{\mathrm{ob}}\|_{P_0}^{2},$$

which softens the initial-state equality of standard MPC into a probabilistic observation factor (Yang et al., 2024). That is a structurally important departure from estimator-then-controller pipelines.

The secure communication framework (Li et al., 2024) and the movable-antenna framework (Yu et al., 14 Aug 2025) extend prediction into communication design. The former uses finite-horizon MPC to repeatedly optimize UAV motion while jointly alternating over beamforming and artificial-noise variables to maximize secrecy rate subject to motion and power constraints (Li et al., 2024). The latter predicts future communication-optimal movable-antenna positions from historical optimal positions, using a Transformer-LSTM trained against PSO-generated secrecy-rate-optimal labels (Yu et al., 14 Aug 2025). The paper explicitly frames this as predictive feedforward positioning rather than a full closed-loop MPC law.

4. Perception, Localization, Communication, and Shared World Models

Collaborative prediction depends on what the team can estimate and communicate. The literature supplies several complementary modules for this layer.

Collaborative localization for micro aerial vehicles provides a decentralized, distributed vision-based estimator in which MAVs with forward-facing monocular cameras combine intra-MAV pose estimation against a shared sparse map with inter-MAV relative pose estimation and covariance-intersection fusion (Vemprala et al., 2019). The state is six-degree-of-freedom pose, map creation begins from inter-MAV feature matching and essential-matrix estimation, and relative corrections are invoked when self-localization quality degrades. For predictive collaborative control, the main value is not trajectory generation but consistent self and relative state estimates with uncertainty-aware fusion (Vemprala et al., 2019).

A second front-end is mobility prediction and profiling. The joint mobility prediction and object profiling method uses a linear kinematic state-space model with unknown acceleration input, then estimates maneuverability profiles through a spike-and-slab-like acceleration model, EM fitting, Bayesian class inference, and online class adaptation (Peng et al., 2018). It predicts future neighboring-object positions and classifies them into maneuverability groups without prior knowledge of all classes. The paper is explicit that this is a prediction-and-profiling module rather than a control law, but its outputs are directly suitable for neighbor modeling, future topology prediction, and heterogeneity-aware coordination (Peng et al., 2018).

Collaborative perception is addressed most directly by MCOP. Its semantic occupancy pipeline uses a Spatial-Aware Feature Encoder, Altitude-Aware Feature Reduction, Dual-Mask Perceptual Guidance, and Cross-Agent Feature Integration to infer 3D semantic occupancy maps from multi-UAV RGB views (Lin et al., 14 Oct 2025). Quantitatively, the paper reports that collaborative occupancy prediction reaches 46.41 mIoU on Air-to-Pred-Occ with only 0.23 MB communication, compared with 41.96 mIoU and 19.14 MB for collaborative PanoOcc, and that AAR reduces transmission from 76.56 MB for uncompressed 3D features to 1.19 MB with only a 0.75 mIoU drop (Lin et al., 14 Oct 2025). This is not future occupancy forecasting, but it provides a communication-efficient 3D semantic world model for downstream prediction and control.

In noisy collaborative sensing, the COMA-based path-planning framework adds an attention-based communication fusion module, SenDFuse, and a CBAM-enhanced actor-critic. The communication model explicitly corrupts exchanged imagery using multiplicative attenuation and additive Gaussian noise, and the denoising-fusion block is pre-trained before being embedded in the MARL pipeline (Zhao et al., 4 Mar 2025). The result is a practical treatment of degraded inter-UAV information exchange rather than idealized perfect communication.

Intent prediction for dynamic obstacles is handled in a separate single-UAV framework that couples tracking, a discrete intent-state model over $$\{\text{forward}, \text{left}, \text{right}, \text{stop}\}$$, trajectory-level prediction for all intents, and MPC-based motion generation (Xu et al., 2024). The paper itself notes that the “MDP” is closer to a Markov chain over intent states than a full control-theoretic MDP. Still, the architectural sequence—tracking, intent belief update, multi-hypothesis trajectory generation, and scenario-conditioned planning—is directly reusable in collaborative UAV systems (Xu et al., 2024).

5. Reconfiguration, Safety, and Robustness

Predictive collaboration in the cited literature is not only about selecting future controls; it is also about maintaining team feasibility under failures, communication change, disturbance, and safety constraints.

CCDS makes reconfiguration a primary control primitive. Network reconfiguration can be triggered when CCDS identifies a demand or when the upper-layer application proactively triggers the network. The framework defines environment-driven adaptation, on-demand adaptation, and subnet reconfigurable submodules, together with two explicit strategies: plan-based reconfiguration and resource pool-based reconfiguration (Jia et al., 2024). It also introduces an Adaptive Communication Strategy (ACS) that adjusts communication links according to task priority, data volume, and network congestion. These mechanisms are supervisory and event-triggered rather than given as formal optimization laws, but they are unusually explicit about how a UAV swarm can merge, split, borrow resources, or reprioritize communication (Jia et al., 2024).

Safety and stability are formalized more tightly in the path-following framework that combines nonlinear model predictive contouring control with an exponentially stabilizing control Lyapunov function and exponential higher-order control barrier functions (Guevara et al., 2024). The path-following Lyapunov function is built from contour and lag errors, while obstacle avoidance is encoded by a relative-degree-two barrier based on distance to obstacles. The resulting optimization simultaneously minimizes contouring error, lag error, control effort, and quaternion attitude error, while enforcing CLF and ECBF constraints (Guevara et al., 2024). Although single-UAV, it is an agent-level safety module that can plausibly be lifted to pairwise UAV-UAV barrier constraints.

The maritime collaborative MPC explicitly enforces inter-UAV separation through predicted-distance constraints and demonstrates that minimum separation is never violated in the reported experiments (Anastasiou et al., 25 Apr 2025). By contrast, the COMA-based informative path-planning paper is explicit that collision avoidance is absent from its formulation, and the intent-aware navigation framework handles safety through inflated obstacle envelopes and hard geometric constraints without formal chance constraints or recursive-feasibility guarantees (Zhao et al., 4 Mar 2025, Xu et al., 2024). A common theme is therefore methodological heterogeneity: some systems encode safety as hard barrier or separation constraints, whereas others rely on architectural robustness, noise-resilient communication, or frequent replanning.

The communication-driven movable-antenna framework highlights a different robustness issue: velocity mismatch between UAV mobility and mechanical antenna repositioning. Its predictive answer is to estimate future optimal antenna positions so that physical reconfiguration begins before the communication geometry becomes stale (Yu et al., 14 Aug 2025). This is not swarm safety, but it is still robustness to latency in a coupled UAV subsystem.

6. Applications, Evidence, and Persistent Open Limits

The application range in the cited works is broad. It includes intelligent agriculture, environmental monitoring, infrastructure inspection, and disaster response at the architectural level (Jia et al., 2024); remote sensing and information collection in 3D belief-space path planning (Zhao et al., 4 Mar 2025); leader-following, landing on a moving base, orbit flight, and dynamic rings crossing (Zhang et al., 2023); maritime search-and-rescue tracking of drifting castaways (Anastasiou et al., 25 Apr 2025); secure UAV communications with beamforming and artificial noise (Li et al., 2024); movable-antenna secure communication in urban environments (Yu et al., 14 Aug 2025); construction-site navigation among dynamic human workers (Xu et al., 2024); and high-level multi-UAV collaborative task planning for target assignment, area search, and patrol (Zhang et al., 30 Jun 2026).

The empirical evidence is correspondingly diverse.

Evidence source	Reported result	Scope
COMA-based IPP	In Env1 at 100%, F1 $0.8432 \pm 0.0260$, entropy $0.2121 \pm 0.0508$; in Env2 at 100%, F1 $u^t\in U^N$0, entropy $u^t\in U^N$1	Information-gathering MARL under noisy communication
CoNi-MPC	Mean iteration time 4.58 ms, standard deviation 1.2 ms, control frequency $u^t\in U^N$2 Hz; fixed-point and fixed-plan tracking errors between 0.11 m and 0.24 m in listed settings	Relative-motion nonlinear MPC
Maritime collaborative MPC	With two agents, average covariance reduced by 87%; with three agents, an additional 10% decrease; average RMSE $u^t\in U^N$3 m in the detailed $u^t\in U^N$4 scenario	Estimation-aware receding-horizon collaboration
MCOP	46.41 mIoU at 0.23 MB on Air-to-Pred-Occ; 47.89 mIoU at 0.23 MB on UAV3D-Occ	Collaborative semantic occupancy prediction
Movable-antenna prediction	At least 49% NMSE reduction and at least 14.76% accuracy improvement at 60 time slots; inference time 8.67 ms	Predictive MA positioning for secure communication
MultiUAV-Plat / Agent4Drone	57.9% task pass rate, 74.6% average task check pass rate, 72.0% global check pass rate	High-level multi-UAV collaborative task planning

These results also clarify several common misconceptions. First, predictive collaborative control is not synonymous with model predictive control. The cited literature includes explicit nonlinear MPC, NMIP-based covariance-minimizing planning, factor-graph joined estimation-and-control, and MARL methods that are predictive only through value estimation and finite mission budgets (Zhang et al., 2023, Anastasiou et al., 25 Apr 2025, Yang et al., 2024, Zhao et al., 4 Mar 2025). Second, not every framework called collaborative control is algorithmically complete. CCDS is explicitly architectural and conceptual rather than a fully specified predictive controller, and Agent4Drone is mission-layer planning rather than low-level control (Jia et al., 2024, Zhang et al., 30 Jun 2026). Third, the nominal smart-agriculture EIA-SEC record does not substantiate its title with usable technical content (Zhou et al., 21 Dec 2025).

Persistent limitations are equally consistent across the literature. Many works lack explicit scalability analysis beyond small teams; several omit formal collision avoidance, bandwidth scheduling, or safety certificates; and many rely on strong assumptions about localization, synchronization, or communication (Zhao et al., 4 Mar 2025, Vemprala et al., 2019, Lin et al., 14 Oct 2025). The factor-graph JPCM is compelling for unifying estimation and control, but its multi-agent extension is not provided (Yang et al., 2024). The secure communication framework gives a strong single-UAV template for motion-resource co-optimization, but no true multi-UAV collaboration (Li et al., 2024). The movable-antenna framework provides a predictive positioning loop, but not a full joint motion-control law over both UAV and antenna plant (Yu et al., 14 Aug 2025).

Taken together, the cited works define predictive MA-UAV collaborative control not as a single established algorithm, but as a family of architectures in which future sensing quality, estimator covariance, map entropy, relative motion, secrecy rate, or mission validation is anticipated and used to coordinate UAV actions. The strongest technical trend is the gradual coupling of layers that were traditionally separated: perception with communication, estimation with control, relative-state prediction with receding-horizon planning, and mission planning with verification and replanning (Yang et al., 2024, Anastasiou et al., 25 Apr 2025, Zhang et al., 30 Jun 2026).