Aerospace autonomy is increasingly defined by the ability to maintain dependable performance when models, measurements, actuators, and communication networks are imperfect. Autonomous aircraft, spacecraft, and cooperative unmanned systems must accommodate nonlinear dynamics, environmental disturbances, and changing mission conditions within limited computational and physical resources. Learning offers valuable mechanisms for adapting to these conditions, but improved prediction or tracking accuracy alone does not establish closed-loop reliability. The central challenge for intelligent robust control is therefore to connect the flexibility of learning with explicit assumptions, stability guarantees, and operational constraints. This Special Issue examines that connection across several aerospace applications and motivates a research agenda in which intelligence and robustness are developed together.
The interaction between control theory and learning provides a useful starting point. Learning-based model predictive control (MPC) can incorporate information from data into prediction and optimization while retaining a framework for addressing constraints [1]. Predictive safety filters offer a complementary architecture in which proposed learning actions are assessed and, when necessary, modified before implementation [2]. These approaches suggest that the role of intelligence should be specified within the control architecture: learning may improve a model, estimate uncertainty, optimize feedback parameters, or generate candidate actions, with the associated guarantees established for the resulting closed loop.
Earlier research on cooperative aerial vehicles illustrates why this architectural perspective matters. Liu et al. [3] developed distributed robust formation control for multiple quadrotors with underactuation, nonlinear coupling, and disturbances in both translational and rotational dynamics. The controller coordinates position and attitude control, with boundedness and eventual tracking within a prescribed neighborhood established under the stated conditions. Liu et al. [4] extended robust fault-tolerant formation control to tail-sitters undergoing aggressive flight-mode transitions, without requiring online identification of actuator faults. Their analysis establishes finite-time entry into specified error neighborhoods. These studies show that robust coordination must account for the interaction between vehicle dynamics, disturbance compensation, and information exchange.
Data-driven control extends this foundation by reducing dependence on detailed dynamic models. Zhao et al. [5] combined a distributed leader-state observer, reinforcement learning, and adaptive fault compensation for attitude synchronization of nonlinear quadrotors. An optimal policy is learned for the nominal augmented system and then incorporated into a fault-tolerant controller; nominal optimality and tolerance to actuator faults are thus distinct properties. Zhao et al. [6] subsequently developed a cascade framework for optimal formation control of heterogeneous quadrotors with unknown dynamics, learning position and attitude policies from system data. Both studies provide theoretical analysis and simulation evidence. Their broader significance lies in connecting learning to a structured cooperative controller while preserving the assumptions needed for policy learning and stability analysis.
Against this background, the five Special Issue contributions discussed here examine complementary aspects of intelligent robust control: prediction informed by physics, adaptive handling of uncertain dynamics, disturbance observation with constrained control, reinforcement learning for distributed coordination, and cooperation under uncertain communication. Their applications differ, but they share a practical question: how can useful adaptation be introduced without obscuring the conditions under which reliable performance can be maintained?
The first direction concerns the use of physical knowledge in learning. Physics-informed machine learning combines observations with governing relationships, offering a way to structure inference when data are limited or the underlying processes are only partially understood [7]. Bao et al. [8] apply this idea to drogue-position prediction for autonomous aerial refueling. Their framework integrates physical information with an attention-augmented long short-term memory network to represent the complex motion of the refueling drogue. This contribution illustrates the value of combining a physical description with a learned predictor for a demanding aerospace task. It also motivates a further question for control design: how should prediction errors be quantified and propagated into guidance or constraint margins? Physical consistency is valuable, but closed-loop guarantees require an explicit treatment of the remaining prediction uncertainty.
The second direction is adaptive robustness for nonlinear flight dynamics. Sun et al. [9] develop robust adaptive trajectory tracking for fixed-wing unmanned aerial vehicles, combining backstepping with adaptation to unknown, state-dependent uncertainty bounds. Their treatment of the kinematic and dynamic loops addresses the coupling between vehicle motion and uncertain dynamics. The contribution emphasizes that useful adaptation can be achieved within an analytical control framework, without making a learned policy responsible for the entire control task. More broadly, adaptive estimation of uncertainty should be judged by the resulting tracking guarantees and required assumptions, as well as by reductions in prior model knowledge. This remains particularly relevant when aerodynamic variations and disturbances interact with underactuation.
A third direction integrates estimation, disturbance rejection, and constraint handling. Chen et al. [10] propose an event-triggered stubborn composite disturbance observer-based MPC framework for spacecraft systems. A radial basis function neural network estimates unknown nonlinearities, while the observer addresses external disturbances and measurement outliers. The MPC controller incorporates input constraints, and sufficient conditions are derived for uniform boundedness of the closed loop. Numerical simulations support the proposed design. The work draws attention to a consequential interface: an estimator that is robust to corrupted measurements must also provide information suitable for constrained prediction and control. Extending such interfaces to quantified estimation errors and feasible control decisions offers a useful route toward more dependable autonomous spacecraft operations.
Communication is a further part of this interface. Distributed event-triggered control and dynamic event-triggered mechanisms provide established foundations for reducing unnecessary information exchanges while retaining coordination properties under specified network conditions [11,12]. In aerospace applications, the choice of an update rule also affects estimation accuracy, processor load, and the information available to each controller. A useful research objective is therefore to assess communication savings together with tracking performance and robustness, including the conditions that prevent excessively frequent transmissions. Communication efficiency becomes more meaningful when its consequences for the physical control system are explicit.
Quan et al. [13] examine output consensus of heterogeneous multi-agent systems through a distributed internal model and event-triggered communication, with reinforcement learning based on centralized training and decentralized execution. The learning component adaptively optimizes local feedback gains to improve transient tracking performance. Its role is specifically separated from the event-triggering mechanism: the reinforcement learning module does not modify the triggering rule or its frequency. This distinction clarifies how the framework combines learning and communication efficiency. The contribution also suggests a useful principle for distributed aerospace control: retain a coordination structure whose behavior can be analyzed, and introduce learning at a clearly defined point within it. The admissible gain choices and the implications of adaptation for stability remain central to that principle.
Zhang et al. [14] address cooperation in a heterogeneous air-ground system comprising tail-sitter unmanned aerial vehicles and ground vehicles. Their robust time-varying pinning controller coordinates the agents through a virtual leader while accounting for switching topologies and uncertain communication links, alongside nonlinear dynamics and other uncertainties. Pinning allows selected agents to receive leader information directly and the remaining agents to coordinate through the network. Viewed alongside the earlier work on tail-sitter formation control [4], this contribution broadens the focus from the robustness of individual vehicle controllers to the robustness of the cooperative structure. Network connectivity and the availability of reference information become part of the conditions governing achievable performance.
Together, these contributions suggest that intelligent robust aerospace control is progressing through closer integration of modeling, inference, feedback, and coordination. They also show why different guarantees should be stated precisely. Bounded tracking errors, asymptotic convergence, optimality for a nominal model, and satisfaction of safety constraints answer different questions. For example, entry into an arbitrarily specified error neighborhood in [3,4] should not be interpreted as exact finite-time convergence to the origin. Likewise, learning an optimal nominal policy in [5] does not by itself establish optimality under actuator faults. Keeping these distinctions explicit makes it easier to identify what an intelligent controller achieves and what additional analysis is required for a particular mission.
A priority for future research is to connect adaptation with explicit safety constraints. Predictive safety filters [2] and control barrier function-based quadratic programs [15] provide established mechanisms for evaluating or restricting candidate actions. Their integration with adaptive or learned components should address the uncertainty in the safety-relevant model and measurements, the feasibility of the constrained control problem, and the effect of limited actuator authority. The resulting guarantee must be matched to these assumptions. This perspective is particularly useful for aerospace tasks in which successful tracking must coexist with collision avoidance, flight-envelope limits, and constraints on control effort.
Another priority is to establish how information quality and vehicle capability jointly limit cooperative performance. Disturbance estimation, fault compensation, and network coordination are often designed separately, yet a loss of actuator effectiveness can reduce the control authority needed to reject disturbances, while communication interruptions can degrade the reference estimates used by neighboring vehicles. Future work should examine these interactions directly and distinguish achievable task performance from desirable nominal objectives. For learning-based controllers, this includes identifying when available data support an update and when uncertainty or degraded information warrants retaining a validated feedback law. Such questions extend the structured learning and robustness principles illustrated by the studies discussed here.
Engineering validation must develop alongside the theory. Representative assessments should examine measurement corruption, disturbances, actuator degradation, and communication changes, with reporting that separates nominal improvements from performance under uncertainty. Progress from simulation to hardware-in-the-loop evaluation and flight experiments would provide stronger evidence of computational feasibility and operational reliability. In June 2026, the European Union Aviation Safety Agency (EASA) released Proposed Issue 03 of its Artificial Intelligence Concept Paper for consultation, extending the discussion to reinforcement learning and advanced automation [16]. As a proposal, it signals an evolving direction for assurance rather than a completed certification framework. For researchers, the relevant implication is to connect performance claims with traceable assumptions and reproducible evidence.
The Special Issue offers perspectives on this connection across aerial refueling, unmanned flight, spacecraft control, and heterogeneous cooperation. Its contributions support a view of intelligent robust control in which learning improves adaptation within architectures that make the conditions for reliable behavior explicit. Continued progress will depend on relating prediction and policy improvements to the guarantees required by the aerospace mission and on testing those guarantees under representative operating conditions. We hope that these studies encourage further work linking advances in learning and control theory to dependable aerospace autonomy.
Conflicts of Interest
The authors served as Guest Editors of this Special Issue, and declare no conflicts of interest.
References
- Hewing, L.; Wabersich, K.P.; Menner, M.; Zeilinger, M.N. Learning-based model predictive control: Toward safe learning in control. Annu. Rev. Control Robot. Auton. Syst. 2020, 3, 269–296. [Google Scholar] [CrossRef] [Scilit]
- Wabersich, K.P.; Zeilinger, M.N. A predictive safety filter for learning-based control of constrained nonlinear dynamical systems. Automatica 2021, 129, 109597. [Google Scholar] [CrossRef] [Scilit]
- Liu, H.; Ma, T.; Lewis, F.L.; Wan, Y. Robust formation control for multiple quadrotors with nonlinearities and disturbances. IEEE Trans. Cybern. 2020, 50, 1362–1371. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Liu, D.; Liu, H.; Lewis, F.L.; Wan, Y. Robust fault-tolerant formation control for tail-sitters in aggressive flight mode transitions. IEEE Trans. Ind. Inform. 2020, 16, 299–308. [Google Scholar] [CrossRef] [Scilit]
- Zhao, W.; Liu, H.; Lewis, F.L. Data-driven fault-tolerant control for attitude synchronization of nonlinear quadrotors. IEEE Trans. Autom. Control 2021, 66, 5584–5591. [Google Scholar] [CrossRef] [Scilit]
- Zhao, W.; Liu, H.; Lewis, F.L.; Wang, X. Data-driven optimal formation control for quadrotor team with unknown dynamics. IEEE Trans. Cybern. 2022, 52, 7889–7898. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Karniadakis, G.E.; Kevrekidis, I.G.; Lu, L.; Perdikaris, P.; Wang, S.; Yang, L. Physics-informed machine learning. Nat. Rev. Phys. 2021, 3, 422–440. [Google Scholar] [CrossRef] [Scilit]
- Bao, X.; Li, Y.; Wang, Z. Prediction of the drogue position in autonomous aerial refueling based on a physics-informed neural network. Aerospace 2025, 12, 540. [Google Scholar] [CrossRef] [Scilit]
- Sun, Y.; Huang, D.; Shi, Z.; Zhong, Y. Robust adaptive trajectory tracking control for fixed-wing unmanned aerial vehicles. Aerospace 2025, 12, 980. [Google Scholar] [CrossRef] [Scilit]
- Chen, J.; Liu, L.; Xu, Y.; Yu, Y. Stubborn composite disturbance observer-based MPC for spacecraft systems: An event-triggered approach. Aerospace 2025, 12, 1010. [Google Scholar] [CrossRef] [Scilit]
- Dimarogonas, D.V.; Frazzoli, E.; Johansson, K.H. Distributed event-triggered control for multi-agent systems. IEEE Trans. Autom. Control 2012, 57, 1291–1297. [Google Scholar] [CrossRef] [Scilit]
- Yi, X.; Liu, K.; Dimarogonas, D.V.; Johansson, K.H. Dynamic event-triggered and self-triggered control for multi-agent systems. IEEE Trans. Autom. Control 2019, 64, 3300–3307. [Google Scholar] [CrossRef] [Scilit]
- Quan, M.; Lan, B.; Long, S.; Wang, Y.; Sun, S. A distributed reinforcement learning method for output consensus of heterogeneous multi-agent systems with event-triggered mechanisms. Aerospace 2026, 13, 487. [Google Scholar] [CrossRef] [Scilit]
- Zhang, J.; Zhao, Z.; Liu, D. Robust time-varying pinning cooperative control for heterogeneous air-ground system with switching topologies and uncertain communication links. Aerospace 2026, 13, 742. [Google Scholar] [CrossRef] [Scilit]
- Ames, A.D.; Xu, X.; Grizzle, J.W.; Tabuada, P. Control barrier function based quadratic programs for safety critical systems. IEEE Trans. Autom. Control 2017, 62, 3861–3876. [Google Scholar] [CrossRef] [Scilit]
- European Union Aviation Safety Agency. EASA Releases Latest Issue of Its Concept Paper on Artificial Intelligence for Comment. Available online: https://www.easa.europa.eu/en/newsroom-and-events/news/easa-releases-latest-issue-its-concept-paper-artificial-intelligence (accessed on 8 October 2026).
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.