cs.ROApr 7, 2026
SaveMARS-Dragonfly: Agile and Robust Flight Control of Modular Aerial Robot Systems
Organizations: Department of Electrical and Computer Engineering, National University of Singapore, Singapore 117583, Singapore · Engineering Design and Innovation Centre, National University of Singapore, Singapore 117583, Singapore
Abstract
Modular Aerial Robot Systems (MARS) comprise multiple drone units with reconfigurable connected formations, providing high adaptability to diverse mission scenarios, fault conditions, and payload capacities. However, existing control algorithms for MARS rely on simplified quasi-static models and rule-based allocation, which generate discontinuous and unbounded motor commands. This leads to attitude error accumulation as the number of drone units scales, ultimately causing severe oscillations during docking, separation, and waypoint tracking. To address these limitations, we first design a compact mechanical system that enables passive docking, detection-free passive locking, and magnetic-assisted separation using a single micro servo. Second, we introduce a force-torque-equivalent and polytope-constraint virtual quadrotor that explicitly models feasible wrench sets. Together, these abstractions capture the full MARS dynamics and enable existing quadrotor controllers to be applied across different configurations. We further optimize the yaw angle that maximizes control authority to enhance agility. Third, building on this abstraction, we design a two-stage predictive-allocation pipeline: a constrained predictive tracker computes virtual inputs while respecting force/torque bounds, and a dynamic allocator maps these inputs to individual modules with balanced objectives to produce smooth, trackable motor commands. Simulations across over 10 configurations and real-world experiments demonstrate stable docking, locking, and separation, as well as effective control performance. To our knowledge, this is the first real-world demonstration of MARS achieving agile flight and transport with 40 deg peak pitch while maintaining an average position error of 0.0896 m. The video is available at: https://youtu.be/yqjccrIpz5o
Figures & tables
Fig. 1: Overview of the proposed MARS-Dragonfly framework. (a) Dragonfly-inspired MARS design. (b) A virtual quadrotor abstraction for arbitrary MARS configurations. (c) Enabling various MARS control tasks with a unified APC controller.
| Mechanism | Mode | Docking Accuracy Requirement | Diagram | Docking or/and Separation Strategy | Actuators ( a set ) | Load |
| ModQuad [ 12 ] | Docking-only | High ( magnet size) | Magnetic attraction | 0 | Low | |
| ModQuad-Vi [ 13 ] | Docking-only | High ( magnet size) | Magnetic attraction | 0 | Low | |
| Magnet matrix [ 14 ] | Separation-only | - | Separation torque due to attitude change | 0 | Low | |
| Electromagnetic [ 29 ] | Separation-only | - | Switch the electromagnet on or off | 4 | Medium | |
| BEATLE [ 26 ] | Integrated | Medium | Docking : magnetic attraction Locking : servo-actuated slider-crank with stick insertion. Separation : a servo reverses the polarity of the magnet, and a servo-driven slider- crank mechanism retracts locking stick. | 2 | High | |
| Our Design | Integrated | Low | Docking : magnetic attraction Locking : asymmetric magnet-actuated twist-lock for passive mechanical interlocking Separation : a servo-driven magnetically repulsive separation | 1 | High |
TABLE I: MARS docking-locking-separation mechanisms
| Method | Model | Controller | Task | Control Performance |
| ModQuad [ 12 ] | Simplified model | PID + Rule-based control allocation | Docking, hovering | Severe oscillation |
| ModQuad-Vi [ 13 ] | Simplified model | PID + Rule-based control allocation | Docking, hovering | Severe oscillation |
| ModQuad-DoF [ 27 ] | Simplified model | PID + Rule-based control allocation | Yaw control, hovering | Medium oscillation |
| Modquad-Gripper [ 28 ] | Simplified model | PID + Rule-based control allocation | Waypoint tracking Object transport | Medium oscillation |
| Ref. [ 14 ] | Simplified model | Torque peak generation | Separation, hovering | Severe oscillation |
| Ours | Virtual quadrotor (Nonlinear model) | Abstracted predictive control + Balanced dynamic control allocation | Docking, Separation Trajectory tracking Agile flight Collision-free navigation Object dragging Agile transport | Smooth and Stable |
TABLE II: MARS model and control
Fig. 2: Docking, locking and separation sequence inspired by dragonfly mating behavior. The process replicates sequential biological actions, including flying \scriptsize1⃝, curling the abdomen \scriptsize2⃝, interlocking together \scriptsize3⃝, and completing mating \scriptsize4⃝, through staged magnetic attraction docking, passive locking, and controlled magnetic repulsion for active separation.
Fig. 3: Microstructural connections in dragonfly mating that inspire the design of macroscopic mechanical structures. (a) The connection of dragonflies includes two docking parts. A1: the male grasps the female’s neck. A2: the secondary genitalia of the male engages with the terminal end of the female’s abdomen for copulation. (b) Microscopic structure of two docking parts. B1: The dense protrusions provide opposing frictional surfaces to support the neck during mating. B2: The microstructure of the female genital cuticle and tile-like scales of the male copulatory apparatus. (c) C1: To mimic the biomechanical control, we employ magnets to simulate a stable connection. C2: The epidermal structure (blue) and tile-like scale surface (yellow) provide enhanced stability. Inspired by this, we design a locking interface with a similar stabilizing effect. (d) The docking mechanism design with magnet array and locking interface.
Fig. 4: Docking and separation mechanism detail. (a) Exploded view of the mechanism. (b) Front and transparent views of the docking structure. Bottom view illustrating the circular gear arrangement that enables controlled separation. (c) Sequential process of approaching, alignment, docking, passive locking, and gear-actuated separation. (d) Illustration of separation failure modes under different disturbances.
Fig. 5: Magnet arrangement optimization. (a) Original and Halbach array for a row of magnets. (b) The magnets are divided into layers, with each layer consisting of two groups. (c) – denote the baseline, manually designed Halbach array and optimized configuration, respectively. The optimized magnet orientations and their corresponding magnetic field distributions are shown, where red regions indicate higher field intensity. The objective function values on the outer and inner surfaces versus iteration count are also presented. Configurations , , and are highlighted in black, purple, and yellow, respectively.
Fig. 6: Coordinate system representation for a MARS with two units.
Fig. 7: Model abstraction and APC framework for MARS. Localization provides odometry for high-level planning. Model abstraction converts a given MARS configuration into a virtual quadrotor, which can be either an equal-arm model or an unequal-arm model, and can be extended to configurations carrying external loads. The abstracted dynamic model and control constraints are incorporated into an APC, which computes real-time control inputs for the virtual quadrotor. An optimization-based dynamic control allocation then maps the virtual control commands to thrust and torque commands for all drone units.
Fig. 8: Model abstraction for various MARS configurations. (a) Algorithm validation using virtual quadrotors formed by one, two, and four units. (b) Equal-arm virtual quadrotor. MARS with the optimal yaw angle exhibits an improvement in maximum net torque over the baseline configurations. (c) Unequal-arm virtual quadrotor. Computed virtual representations of MARS configurations, illustrating unbalanced torque along the - and -axes. Torque approximation error of each model abstraction, computed using Eq. ( 19 ).
Fig. 9: Model abstraction for arbitrary-shaped transportation using MARS. Unequal- and equal-arm quadrotor formations with (a) 2 units and (b) 3 or 4 units, illustrating object transport.
Fig. 10: Simulation of trajectory tracking and object transport. (a)–(e) Trajectory tracking over ten consecutive laps along a circular path under different configurations. (f)–(j) Circular trajectory tracking on a 1.5 m radius path at 2 m/s with a 1.5 kg payload under different task settings.
| (a) | (b) | (c) | (d) | (e) | Average |
| 0.0261 | 0.0364 | 0.0385 | 0.1264 | 0.1329 | 0.0721 |
TABLE III: Trajectory Tracking Error in Fig. 10
| Config | Method | Vel (m/s) | Payload (kg) | Method Avg | ||
| 0.5 | 1.0 | 1.5 | ||||
| 2 1 | Modquad [ 28 ] | 1 / 2 | – | |||
| Ours: APC | 1 | 0.1349 | 0.1599 | 0.1327 | 0.1815 | |
| 2 | 0.2134 | 0.2338 | 0.2142 | |||
| 3 1 | Modquad [ 28 ] | 1 / 2 | – | |||
| Ours: APC | 1 | 0.1960 | 0.1111 | 0.1436 | 0.1619 | |
TABLE IV: Tracking error (m) under different velocities, payload weights, and configurations.
| Payload (kg) | Method | XY Error (m) | Max Height Error (m) | Stability |
| 0.5 | Modquad [ 12 , 28 ] | N/A | N/A | N/A |
| Ours: APC | 0.2982 | 0.5294 | Stable | |
| 1.0 | Modquad [ 12 , 28 ] | N/A | N/A | N/A |
| Ours: APC | 0.3788 | 2.5023 | Unstable |
TABLE V: Comparison of Agile Transport
| Model + Controller | Allocation | Task | Average | ||
| Tracking | Payload | Dragging | |||
| Simplified model + PID | Rule-based | ||||
| Ours: APC | Rule-based | ||||
| Ours: APC | Dynamic | ||||
TABLE VI: Tracking Error (m) under Different Ablation Settings
Fig. 11: Passive docking and lock sequences. The timer format is min:sec:ms . (a) Before docking: the servo is not powered, and the mechanism is in the standby state. (b) Initial docking: the male and female components begin to engage under the guidance of magnetic attraction. (c) The male docking mechanism is fully inserted into the female component and begins to lock. (d) The male component passively rotates due to the asymmetric magnetic field, reaching the designed locking position. (e) A pendulum test demonstrates the strong mechanical strength of the docked connection.
Fig. 12: Manually triggered active separation rotates the male component to repel the connected module. (a) Initial locking state. (b) At 00:00:00, the male component of the rotational docking mechanism is manually rotated to unlock the docking-separation interface. The repulsive force from embedded magnets assists the separation process. (c) Fully separated state at 00:00:04.
| Method | Docking | Separation | Agile tracking | Agile flight | Collision free | Agile transport |
| Modquad [ 12 ] | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Modquad-Vi [ 13 ] | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Ref [ 14 ] | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ |
| Ref [ 43 ] | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ |
| Ours: APC | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
TABLE VII: Summary of Experimental Tasks Performed by MARS
| Method | Docking Mechanism | Max x/y Acceleration ( ) | Normalized height error |
| Modquad [ 12 ] | Magnet | Not reported | 2.02 = 0.097 / frame size |
| ModQuad-Vi [ 13 ] | Magnet | 4.9 | 5.15 = 0.577 / frame size |
| Ours | Mode shift | 3.47 | 0.28 = 0.042 / frame size |
TABLE VIII: Comparison of docking methods
Fig. 13: Mid-air docking demonstration. The timer format is min:sec:ms . (a) The hover unit maintains a stable position while the docking unit takes off, approaches, and actively adjusts its trajectory to achieve precise mid-air docking. (b) The docking process starts at 00:47:00 and successfully ends at 00:47:19. (c) Time-series plots of acceleration, attitude angles, and position for both units, showing active trajectory correction of the docking unit, and the docking collision, illustrating the stability of the proposed mid-air docking approach.
Fig. 14: Mid-air separation demonstration. The timer format is min:sec:ms . (a) Two drones docked together take off simultaneously, perform a mid-air separation maneuver, and then independently land. (b) After switching to separation mode, the drones successfully disengage from each other during the time interval 00:12:12 to 00:13:40. (c) Experimental data during the demonstration, showing the acceleration along x, y, z axes, attitude angles (pitch, roll, yaw), and position trajectories. The key events is marked on the plot, including take-off, mid-air separation, and landing states.
Fig. 15: Separation comparison experiments. (a) One side enabled with repulsive force; the other disabled. (b) Initial position at 00:09:13. (c) The mechanism employing a repulsive force achieves successful separation at 00:09:25. (d) Without repulsive force, separation fails until 00:11:19, with a delay of approximately 2 s. (e) Successful separation occurs at 00:12:04.
| Method | Separation method | Maximum tilting angle ( ∘ ) | Separation time (s) |
| Ref. [ 14 ] | Separation torque | about 28 | 5 |
| Ref. [ 29 ] | Electromagnetic | 4 (Fully Actuated) | 3 |
| Ours | Mode shift | 7.2 | 2.9 |
| Ours | Mode shift + repulsive force | 5.8 | 0.5 |
TABLE IX: Comparison of separation methods
Fig. 16: Composite images of trajectory tracking. (a) Trajectory tracking at a velocity of 1 m/s . (b) Circular trajectory tracking with the docking mechanism engaged and a desired yaw angle. (c) Collision-free navigation illustrating the robust tracking performance of MARS.
Fig. 17: Agile trajectory tracking. (a) Pitch-angle responses along agile trajectories at , , and without retuning the APC parameters, demonstrating robustness to varying attitude demands. (b) Pitch-angle tracking in (a) was evaluated at the planned spatial path, further demonstrating the robustness of the agile flight controller. (c) Roll-angle tracking along an agile trajectory. (d) The MARS maintains speeds exceeding while following a -radius circular path; the roll and pitch angles approach , indicating high-agility performance.
Fig. 18: Object transport and dragging (600 g payload, 20% of MARS weight). (a) Circular tracking at 2 m/s demonstrates robust performance while carrying the payload. (b) Payload dragging on the ground further demonstrates robustness under payload–ground contact disturbances. (c) Overlaid trajectories highlight accurate tracking performance.
Fig. 19: Object transportation with a sloshing load. (a) Composite images show the basketball’s oscillatory motion during transport. (b) Attitude error reflects sloshing effects; position tracking confirms controller robustness.
Fig. 20: Agile transport carrying 100 g, 200 g, and 600 g payloads. (a) Pitch angle responses along agile trajectories at without retuning the APC parameters, demonstrating robustness to varying unmodeled payload disturbances. (b) Overlaid pitch angle trajectories for different payloads at the planned pitch angle and position, further highlighting robust tracking performance of the agile flight controller.
| Task | Payload (kg) | Position Error (m) | Attitude Error ( ∘ ) |
| Tracking | 100 g | 0.0512 | 2.00 |
| 200 g | 0.0409 | 2.23 | |
| 600 g | 0.0340 | 1.77 | |
| Avg | 0.0420 | 2.00 | |
| Dragging | 100 g | 0.0523 | 2.46 |
| 200 g | 0.0492 | 2.51 |
TABLE X: Comparison of Agile Transport
More in this topic
Aerial Robotics
Aerial robotics studies the design, perception, control, and coordination of flying robots for navigation, inspection, transport, and physical interaction.
See all 340 papers