MARS: Modular Aerial Robot Systems

• Modular Aerial Robot Systems (MARS) are aerial platforms that assemble standardized drone units into a single, reconfigurable rigid structure.
• Their modular design integrates propulsion, sensing, and payload modules to enable agile control, enhanced actuation, and improved fault tolerance.
• Recent advances optimize configuration using aerodynamic constraints and collective wrench mapping, ensuring safe, dynamic reassembly and robust performance.

Modular Aerial Robot Systems (MARS) denote aerial platforms whose flight characteristics are determined by assembling, rearranging, or augmenting standardized units rather than by fixing a monolithic airframe once at design time. In the strict rigid-body formulation that now dominates the term, MARS consist of multiple drone units physically bound together into a single flying structure that can self-reconfigure in response to mission requirements or fault conditions; adjacent strands of the literature extend the idea to attachable flight kits for rigid payloads, heterogeneous multirotor modules with task-dependent controllable degrees of freedom, and articulated self-assembling aerial manipulators (Huang et al., 12 Mar 2025, Huang et al., 12 Mar 2025, Huang et al., 17 Sep 2025, Mu et al., 2019, Xu et al., 2022, Sugihara et al., 19 May 2026).

1. Definition and conceptual boundaries

In the core control-and-reconfiguration literature, MARS are defined as assemblies of multiple drone units that behave as one rigid aerial vehicle after docking. One line of work describes MARS as “multiple drone units assembled into a single, integrated rigid flying platform,” while another treats them as multiple small drone units that are physically docked so that the resulting system behaves as one rigid aerial vehicle with shared dynamics, a common center of mass, and a common attitude (Huang et al., 12 Mar 2025, Huang et al., 17 Sep 2025). This definition places the central technical problem at the level of collective wrench generation, controllability, and configuration-dependent dynamics rather than at the level of formation-keeping alone.

A persistent misconception is to equate MARS with any multi-UAV team. The rigid-body literature explicitly separates docked MARS from swarm or multi-quadrotor systems: in a swarm, independent vehicles coordinate trajectories, whereas in MARS all rotors contribute to a common thrust-and-torque vector for one shared dynamics model (Huang et al., 17 Sep 2025). This distinction matters because controllability, allocation, and fault tolerance are then properties of the assembled structure rather than of isolated vehicles.

At the same time, the broader modular-aerial literature shows that the idea of modularity exceeds one single embodiment. UFO turns an arbitrary rigid object into a multirotor by affixing a control module and multiple propelling modules to the payload, which then serves as the airframe (Mu et al., 2019). LEGION instead uses identical flying units with docking interfaces and joints at both ends, so that self-assembly produces an articulated aerial manipulator rather than only a rigid body (Sugihara et al., 19 May 2026). This suggests a wider interpretation in which “modular aerial systems” cover rigid assemblies, articulated assemblies, and deployable flight-capability modules, even when only part of that spectrum is labeled MARS explicitly in the source literature.

2. Structural architectures and hardware modularity

A recurrent architectural theme is the decomposition of the aerial robot into propulsion, sensing, power, and task modules. “Space CoBot” embodies this directly: its modular design comprises a hexrotor propulsion system and a stack of modules including batteries, navigation cameras, a screen for telepresence, a robotic arm, space for extension modules, and a pair of docking ports that can be used both for docking and for mechanically attaching two Space CoBots together (Roque et al., 2016). Because its kinematics are holonomic, translational and rotational components can be fully decoupled, making modular payload addition compatible with human-facing tasks in inhabited microgravity environments.

A second architectural lineage uses homogeneous or heterogeneous multirotor modules as the basic building block. H-ModQuad is built from cuboid modules containing quadrotors with tilted propellers; by assembling different module types, the vehicle can increase its actuated degrees of freedom from 4 to 5 and 6, and the modules differ functionally through their actuation properties rather than through their basic mechanical form factor (Xu et al., 2022). Here modularity is not only structural but also actuation-theoretic: the global allocation matrix changes rank as module orientations and placements change.

A third lineage detaches flight capability from the payload entirely. UFO introduces a control module and propelling modules that can be affixed to an arbitrary rigid object, so that the object itself becomes the multirotor airframe; hovering and trajectory tracking were demonstrated with different dummy payloads weighing approximately 200–800 grams and with four to eight propelling modules (Mu et al., 2019). In this formulation, modularity appears as a reusable flight kit for rigid objects rather than as a predesigned reconfigurable chassis.

At the platform level, the MRS Modular UAV Hardware Platforms work extends modularity to the family-of-vehicles scale. The proposed UAV designs are modular in actuators, sensor configurations, and even UAV frames, supporting standard quads, coaxial multirotors, and custom octocopters while keeping a common electrical and software backbone across indoor, outdoor, single-robot, and multi-robot deployments (Hert et al., 2023). This hardware perspective is less focused on in-flight self-assembly than the strict MARS literature, but it supplies the standardized interfaces and parametric modeling assumptions on which such systems depend.

3. Wrench spaces, actuation capability, and configuration selection

Across the literature, the common mathematical object is a configuration-dependent mapping from actuator inputs to body wrench. In task-centric modular multirotor design, the total wrench is written as

$\mathbf{w} = A\mathbf{u},$

and the feasible wrench set is

$$\mathcal{W} = \left\{ A\mathbf{u} \;\middle|\; 0 \preceq \mathbf{u} \preceq f_{\max},\ \mathbf{u}\in\mathbb{R}^{4n} \right\},$$

which is a convex polytope—specifically a zonotope—in wrench space (Xu et al., 2023). This formulation turns configuration design into a containment problem: a modular aerial structure is adequate for a task if the task wrench set lies inside its feasible wrench set.

H-ModQuad extends this perspective with actuation ellipsoids and configuration-dependent reference frames. By analyzing the singular structure of the force part of the allocation matrix, it defines a body frame that maximizes thrust performance and thereby aligns control with the dominant actuation directions of the assembled structure (Xu et al., 2022). The same work uses polytopes to compare actuation capability against task requirements, making payload support, tilt capability, and the transition from 4 to 5 to 6 controllable degrees of freedom properties of the module arrangement rather than of a fixed platform.

Recent configuration optimization work adds aerodynamic constraints that earlier planar-layout studies often neglected. In “Downwash-aware Configuration Optimization for Modular Aerial Systems,” each module induces a capsule-shaped downwash volume, and feasibility requires pairwise clearance:

$$\| \boldsymbol{d}_i - \boldsymbol{d}_j \|_2^2 \ge 4r^2.$$

The resulting nonlinear program jointly chooses connector angles and rotor inputs while enforcing actuation limits and downwash bounds, and it is embedded in a pipeline that enumerates non-isomorphic connection topologies at scale (Li et al., 20 Feb 2026). This shifts configuration design from purely geometric combinatorics toward aerodynamically constrained synthesis.

“MARS-Dragonfly” pushes the same modeling program into agile flight control by introducing a force-torque-equivalent virtual quadrotor. The physical MARS wrench is written

$\mathbf{u}_V = \mathbf{G}_M \mathbf{f},$

and a virtual quadrotor is constructed whose feasible wrench polytope is contained inside the physical MARS wrench polytope, so that existing quadrotor controllers can be reused across different connected formations (Huang et al., 7 Apr 2026). The same framework optimizes yaw angle to maximize control authority, making orientation itself a configuration variable in the actuation design loop.

4. Estimation, control, and software substrates

Control architectures in modular aerial robotics must absorb large configuration variation without per-configuration redesign. “Space CoBot” addresses this in a holonomic microgravity setting with separate converging controllers for position and attitude, together with realistic-physics simulation that evaluates sensitivity to sensor noise and to unmodeled dynamics such as load transportation (Roque et al., 2016). Even in this early modular platform, the central issue is the decoupling of translational and rotational control under a nonstandard propulsion geometry.

UFO treats configuration uncertainty explicitly. Its IMU-based estimation strategy reconstructs module orientations, relative positions, mass, inertia, and the configuration matrix from distributed IMU measurements, and an adaptive geometric controller then refines uncertain parameters online; the reported result is hovering and trajectory tracking with position errors of a few centimeters after parameter convergence (Mu et al., 2019). This is a particularly clear example of MARS control as estimation-plus-adaptation rather than as fixed-model regulation.

H-ModQuad contributes a complementary result at the actuation-rank level by proposing three controllers, one each for 4, 5, and 6 controllable degrees of freedom, and validating the independence of translation and orientation experimentally when full or partial actuation is available (Xu et al., 2021). The implication is that modularity changes not just vehicle shape but the very structure of the admissible feedback law.

On the software side, Aerostack2 provides a layered, ROS 2 based framework for multi-robot aerial systems with standardized inter-process communication, platform and sensor abstractions, plugin-managed basic robotics functions, behavior libraries, behavior-tree mission control, and application-level integrations (Fernandez-Cortizas et al., 2023). Its relevance to MARS lies in the software realization of modularity: controllers, estimators, behaviors, mission logic, and platform interfaces can be assembled and replaced in a way that mirrors the reconfigurability of the hardware.

5. Reconfiguration, docking, fault tolerance, and articulated morphology

Fault tolerance is one of the strongest motivations for MARS. “Robust Fault-Tolerant Control and Agile Trajectory Planning for Modular Aerial Robotic Systems” models MARS as rigid assemblies of multiple quadrotor units and proposes a control reallocation method that redistributes the collective force and torque to individual units according to their moment arm relative to the center of MARS mass, together with a collision-avoiding and dynamically feasible trajectory planner for arbitrary configurations (Huang et al., 12 Mar 2025). The paper explicitly targets oscillations during docking and separation, making reallocation smoothness a first-class control objective rather than a secondary implementation detail.

A more structural formulation appears in “Robust Self-Reconfiguration for Fault-Tolerant Control of Modular Aerial Robot Systems,” where practical controllability under actuator limits is summarized by the controllability margin

$$\mathrm{CM} := \zeta(\mathbf{g}, \partial\boldsymbol{\Omega}),$$

the signed distance from the gravity wrench to the boundary of the admissible wrench set (Huang et al., 12 Mar 2025). Reconfiguration is then posed not merely as finding a good final structure, but as finding disassembly and assembly sequences whose intermediate subassemblies remain controllable. This directly corrects an earlier limitation of self-reconfiguration work: a final shape may be attractive while intermediate shapes are dynamically unsafe.

TransforMARS generalizes this controllability-centered view to arbitrary shapes and multiple rotor and unit faults. It introduces virtual minimum controllable subassemblies containing faulty units, plans feasible disassembly-assembly transport sequences for units or subassemblies, and ensures continuous in-air stability throughout the process (Huang et al., 17 Sep 2025). In contrast with rectangular-grid assumptions, hollow and irregular shapes such as heart and letter-like formations become admissible reconfiguration targets.

Docking mechanics and control have also become substantially more sophisticated. MARS-Dragonfly combines passive docking, detection-free passive locking, and magnetic-assisted separation using a single micro servo with a two-stage predictive-allocation pipeline: a constrained predictive tracker computes virtual inputs under force/torque bounds, and a dynamic allocator maps them to module-level commands with balanced objectives to keep motor commands smooth and trackable (Huang et al., 7 Apr 2026). This directly addresses a practical failure mode of earlier rule-based allocation schemes, namely discontinuous and effectively unbounded commands during docking, separation, and aggressive tracking.

LEGION extends reconfiguration from rigid-body assembly to articulated morphology. Each unit is a fully actuated tri-rotor with docking interfaces and joints at both ends; multiple units autonomously dock in flight, maintain a zero-clearance interlock by controlling contact force and torque, and then execute pushing, pulling, rotating, grasping, and carrying as an articulated flying manipulator (Sugihara et al., 19 May 2026). Here modularity is no longer only about changing actuation authority or fault recovery; it becomes a means of switching between nimble individual flight and collective physical interaction.

6. Applications, validation, and open problems

The application space of MARS is unusually broad because the same modular principles recur in distinct environments. Space CoBot targets telepresence and cooperative mobile manipulation in inhabited microgravity environments such as orbiting space stations (Roque et al., 2016). The MRS hardware platforms emphasize rapid transition from simulation and laboratory setups to uncertain indoor and outdoor conditions through interchangeable actuators, frames, sensors, and software components (Hert et al., 2023). Recent fault-tolerant MARS work explicitly identifies search and rescue, payload transport, and inspection as natural use cases for self-reconfigurable aerial structures (Huang et al., 12 Mar 2025).

Experimental validation has also moved beyond quasi-static hovering. MARS-Dragonfly reports simulations across over 10 configurations and real-world experiments showing stable docking, locking, separation, agile transport, and 40 deg peak pitch while maintaining an average position error of 0.0896 m (Huang et al., 7 Apr 2026). LEGION demonstrates autonomous in-flight self-assembly, outdoor morphing, and core manipulation primitives including pushing, pulling, rotating, grasping, and carrying (Sugihara et al., 19 May 2026). These results indicate that modular aerial systems are no longer confined to proof-of-concept assembly maneuvers; they are being evaluated under contact, payload, and configuration-switching regimes.

A broader reading of the field suggests that modularity in aerial robotics also includes system-of-systems compositions rather than only docked composites. The marsupial walking-and-flying exploration system pairs a legged carrier with a selectively deployable aerial robot for collaborative mapping and exploration path planning in unknown environments (Petris et al., 2022). SphereX, in turn, proposes a network of small hopping, flying, or rolling robots for caves, lava tubes, and skylights on Mars, explicitly leveraging multi-robot redundancy, distributed command and control, and parallel task execution (Kalita et al., 2017). This suggests that the conceptual perimeter of MARS overlaps with heterogeneous deployment architectures whenever modular composition changes the reachable task envelope.

The principal open problems are now well defined. Downwash-aware optimization shows that planar layout alone is insufficient once dense 3D assemblies are considered (Li et al., 20 Feb 2026). Fault-tolerant control papers still rely on assumptions such as rigid docking, negligible aerodynamic interaction, simplified dynamics, and perfect communication, and they do not provide formal Lyapunov-style stability guarantees for the complete reconfiguration process (Huang et al., 12 Mar 2025, Huang et al., 12 Mar 2025). Software frameworks such as Aerostack2 reduce fragmentation, but large-scale coordination, richer built-in behaviors, and standardized safety/performance metrics for heterogeneous modular aerial systems remain incomplete (Fernandez-Cortizas et al., 2023). In aggregate, the literature points toward a next phase in which actuation-aware design, aerodynamics, formal controllability, and autonomy middleware are treated as one integrated MARS problem rather than as separate subfields.