Multi-Rotor Micro Aerial Vehicles

• Multi-Rotor MAVs are lightweight rotorcraft with multiple rotors enabling agile, precise 6-DOF flight under tight energy and processing constraints.
• They integrate advanced sensor fusion and state estimation for robust navigation in complex, GPS-denied environments using technologies like LiDAR and RF backscatter.
• Their design supports multi-agent coordination, formation control, and autonomous inspection through sophisticated motion planning and resilient control algorithms.

Micro-Rotor Micro Aerial Vehicles (MAVs) are a class of rotorcraft with multiple lift-producing rotors, typically in quadrotor, hexarotor, or higher order multirotor configurations, that operate at the sub-2 kg scale with on-board computation, sensing, and actuation. Their design prioritizes agile flight, precise hovering, and full 6-DOF maneuverability under severe mass, energy, and processing constraints. MAVs are deployed across domains from autonomous inspection and exploration in GPS-denied environments to collaborative manipulation and active perception. State-of-the-art research encompasses vehicle dynamics, multi-modal sensing, robust state estimation, trajectory planning, resilient control under failure and uncertainty, formation and multi-agent coordination, and real-time perception–action loops.

1. Vehicle Dynamics and Modeling

Rotorcraft MAVs universally employ Newton-Euler rigid-body models, accounting for the interaction between $n$ independently actuated rotors, their collective and distributed thrust, and external disturbances. In an inertial frame with position $p\in\mathbb{R}^3$ and attitude $R \in SO(3)$:

$m\ddot p = m g e_3 - R\sum_{i=1}^n T_i e_3$

$$J\dot{\omega} + \omega \times J \omega = \sum_{i=1}^n \tau_i e_3$$

where $T_i = k_F\Omega_i^2$ and $\tau_i = k_M\Omega_i^2$, with $\Omega_i$ the $i$th rotor speed, $k_F, k_M$ aerodynamic coefficients, and $J$ the inertia tensor (Kulkarni et al., 2024). A standard twelve-state full nonlinear model encompasses position, velocity, rotation, and body rates, with actuator mappings for each rotor. Real-world MAVs layer inner-loop attitude-rate stabilization, often realized as first-order low-latency PID or LQR controllers, over this high-fidelity plant (Nguyen et al., 2020). Advanced platforms such as the fully-actuated MOMAV employ six rotating arms in octahedral symmetry, allowing independent force and moment control via online Sequential Quadratic Programming (SQP) allocation under torque and thrust constraints (Ruggia, 10 Jun 2025).

2. State Estimation and Sensor Fusion

Navigating complex, GPS-denied environments requires tight sensor fusion architectures. Core IMU streams (tri-axial accelerometers, gyroscopes) provide high-rate motion cues, fused with vision (monocular, stereo, event cameras), LiDAR, UWB, RF backscatter, and barometric data within EKF, UKF, or error-state Kalman filter (ESKF) frameworks (Kulkarni et al., 2024, Zhang et al., 2019, Helm et al., 2018). For instance, robust LiDAR–inertial SLAM achieves sub-decimeter mapping and loop-closure via tightly-coupled Fast-LIO2 with VGICP scan matching (Pan et al., 2024). In multipath or visually degraded environments, infrastructure-free RF backscatter tags (Marvel) enable 6-DoF pose estimation at up to 50 m, integrating factor-graph fusion of channel phase/angle/yaw measurements with inertial odometry—achieving 34 cm positioning and 5° yaw accuracy in smoke-filled or GPS-denied conditions (Zhang et al., 2019). UWB ranging, combined with on-board EKF, supports heading-independent relative localization between MAVs, achieving mean localization errors of 0.18–0.52 m over 5 m baselines (Helm et al., 2018).

Collaborative localization schemes exploit inter-agent visual feature sharing, performing distributed global map triangulation, relative-pose estimation via multi-view geometry, and covariance-intersection fusion for consistent group state estimation—empirically achieving trajectory errors below 1% of range (Vemprala et al., 2019).

3. Motion Planning, Coverage, and Formation Control

Trajectory generation on MAVs leverages the system's differential flatness, enabling polynomial spline or motion primitive-based planning in $(x,y,z,\psi)$ via convex or search-based methods. State lattices constructed from flat outputs and controlled with canonical-form dynamics enable A*, LPA* resolution-optimal planning with hard constraints for collision avoidance, field-of-view, motion uncertainty (via APF soft-costs), and dynamic obstacles. In typical scenarios, full-path replanning achieves sub-10 ms runtimes per update and reliably avoids static/moving obstacles as well as inter-agent collisions (Liu et al., 2018).

Coverage planning for inspection missions (e.g., wind turbines) processes 3D point clouds to extract branches, performs region clustering, applies FOV and safety-distance offsetting, and assigns optimized scan paths to each MAV to ensure complete coverage with collision avoidance. Segmentation employs CSF for ground extraction, RANSAC for planar features, and Euclidean clustering for object detection (Kanellakis et al., 2019, Pan et al., 2024). Instance-aware scan path planners deploy spiral or CCPP grids on identified objects, combined with B-spline-based global optimization under collision, smoothness, and dynamics constraints. Instance segmentation accuracy F1≈0.96 and 100% scan-path execution success have been reported for large-scale indoor facilities (Pan et al., 2024).

Multi-agent formation and cooperative transport are enabled by centralized or decentralized Nonlinear Model Predictive Control (NMPC) frameworks. All-vehicle state and input sequences are optimized over realistic 6-DoF dynamics and local relative sensing, with real-time iteration (RTI) solvers delivering centimeter-level 3D formation accuracy and robust collision avoidance (Erunsal et al., 2019, Kamel et al., 2017). In robust collaborative transportation, decentralized admittance control, tuned via $\mu$-analysis, ensures performance and worst-case stability under modeling and estimator uncertainty for up to $N=10$ MAVs, validated with sub-decimeter tracking in experiments (Tagliabue et al., 2017).

4. Control Algorithms and Robustness to Disturbance

Multirotor MAV control stacks are predominantly hierarchical: inner-loop attitude control, outer-loop position/velocity control, and high-level model predictive or adaptive controllers. Model Predictive Control (MPC) is the dominant paradigm for optimizing tracking accuracy under full system dynamics and physical constraints. Linear MPC (solved via LQ/QP) is computationally efficient for small-disturbance or hover, but Nonlinear MPC (NMPC, solved by multiple shooting or sequential QP) maintains accuracy for aggressive maneuvers and large-angle flight—the performance gap exceeding 10 cm RMSE in dynamic benchmarks (Nguyen et al., 2020). State/input constraints, obstacle avoidance, and real-time disturbance estimation through augmented observers are standard (Nguyen et al., 2020, Kulkarni et al., 2024). Fault-tolerant MPC can dynamically reallocate control in the event of rotor failures, updating the control allocation matrix and actuator bounds to enable continued flight after up to three rotor failures (Nguyen et al., 2020).

Adaptive and learning-based controllers are increasingly adopted: Parsimonious Adaptive Controllers (PAC), based on evolving neuro-fuzzy architectures, deliver real-time adaptation with minimal parameter overhead and outperform conventional PID and feed-forward neural-network controllers in RMSE and settling time (Ferdaus et al., 2018). Reinforcement learning integrated with MPC—through value-function approximation, neural-network-based dynamics modeling, or explicit policy distillation—improves constraint handling and generalization to new environments (Nguyen et al., 2020, Singh et al., 8 Apr 2025). RL-based policies trained on depth input achieve 91% reduction in training time and, when transferred to real-world indoor deployments, deliver safe flight lengths exceeding 3x competing RGB-based approaches (Singh et al., 8 Apr 2025).

Energy and environmental disturbance resilience is addressed via explicit disturbance observers (DOBs), which estimate wind or contact forces and enable stable, input-to-state stable flight in up to 10 m/s gusts (Daftry et al., 2016), as well as via hardware (e.g., collision cages, foldable frames) and perception redundancy (multi-modal CompSLAM (Kulkarni et al., 2024)).

5. Multi-Agent Perception and Cooperative Tasks

Active perception—with on-board deep neural networks (e.g., SSD Multibox, VGG backbone) and inter-MAV data fusion—enables robust, real-time cooperative tracking and formation around moving objects. Low-latency inter-robot communication of condensed 3D pose estimates, selective region-of-interest processing, and model-predictive formation control achieves 3D tracking MSE down to 0.1 m² and 90+% detection rates in challenging outdoor fields (Price et al., 2018). For GNSS-/compass-denied or magnetically compromised environments, heading-independent UWB ranging, supplemented by time-delayed trajectory tracking and high-precision clock synchronization, allows leader-follower flight with sub-0.25 m average localization error, tolerant of multi-agent and communication-induced delays (Helm et al., 2018).

Collaborative visual SLAM and distributed map fusion are deployed for large-area inspection and exploration. Each MAV builds local visual maps; frequent relative-pose updates and covariance intersection fusion among team members maintain absolute and drift-free 6-DoF localization, even when individual robots temporarily lose visual features (Vemprala et al., 2019). Sensor fusion schemes accommodate low bandwidth and asynchrony, ensuring robustness to communication outages.

6. Specialized Capabilities and Benchmarks

MAVs increasingly operate in aggressive, energy-constrained, and dynamic settings. Vision-based perching up to 90° inclines is achieved via polynomial-spline planning, unscented KF VIO, geometric $SE(3)$ control, and global constraint verification by Sturm-sequence root counting—enabling reliable high-rate, high-acceleration trajectories and energy-saving perching maneuvers (Mao et al., 2021). Automated landing on fast-moving ground vehicles (up to 50 km/h) leverages multi-sensor fusion (visual, GNSS, IMU in a Kalman filter) and a hybrid proportional navigation/PID controller, delivering 0.2 m RMSE under real-world, high-speed conditions (Borowczyk et al., 2016).

Autonomous inspection frameworks tightly integrate hierarchical perception, B-spline motion planning, neural 3D Gaussian splatting for real-time photo-realistic site reconstruction, and error-state filtering for continuous relocalization. In indoor settings up to 4000 m², these systems demonstrate F1=0.96 in segmentation, mean trajectory tracking errors below 0.1 m, and PSNR > 30 dB for 3D rendering (Pan et al., 2024). Visual-inertial, UWB-fused, or backscatter-based state estimation protocols guarantee operation in environments inaccessible to LiDAR or GPS (Zhang et al., 2019, Kanellakis et al., 2019, Pan et al., 2024).

7. Open-Source Tools, Benchmarks, and Research Frontiers

A strong open-source ecosystem accelerates MAV research. Notable software includes mav_control_rw (ETH-ASL, LMPC and NMPC samples), rpg_mpc (perception-aware NMPC), denmpc (real-time NMPC), and general-purpose NMPC and QP solvers (ACADO, ACADOS, CVXGEN, YALMIP, do-mpc, Control Toolbox) (Nguyen et al., 2020). RL environments and datasets (AirSim, Unreal Engine meta-environments) and end-to-end reference pipelines for active perception, onboard DNN inference, and inter-agent communication are available (Singh et al., 8 Apr 2025, Price et al., 2018).

Current bottlenecks include battery endurance, mapping and exploration scalability, resilient operation under multi-modal sensor failure, and direct physical interaction. Promising research directions emphasize lightweight learned policies for resilience, energy-aware motion and hardware co-design, semantic mapping with real-time object-level fusion, and architectural advances for safe contact, manipulation, and multi-agent coordination in highly dynamic, unstructured settings (Kulkarni et al., 2024).

---

References

• (Nguyen et al., 2020) Model Predictive Control for Micro Aerial Vehicles: A Survey
• (Kulkarni et al., 2024) Aerial Field Robotics
• (Ruggia, 10 Jun 2025) MOMAV: A highly symmetrical fully-actuated multirotor drone using optimizing control allocation
• (Zhang et al., 2019) RF Backscatter-based State Estimation for Micro Aerial Vehicles
• (Helm et al., 2018) On-board Range-based Relative Localization for Micro Aerial Vehicles in indoor Leader-Follower Flight
• (Price et al., 2018) Deep Neural Network-based Cooperative Visual Tracking through Multiple Micro Aerial Vehicles
• (Vemprala et al., 2019) Collaborative Localization for Micro Aerial Vehicles
• (Pan et al., 2024) Developing Smart MAVs for Autonomous Inspection in GPS-denied Constructions
• (Liu et al., 2018) Towards Search-based Motion Planning for Micro Aerial Vehicles
• (Kamel et al., 2017) Nonlinear Model Predictive Control for Multi-Micro Aerial Vehicle Robust Collision Avoidance
• (Erunsal et al., 2019) Nonlinear Model Predictive Control for 3D Formation of Multirotor Micro Aerial Vehicles with Relative Sensing in Local Coordinates
• (Tagliabue et al., 2017) Robust Collaborative Object Transportation Using Multiple MAVs
• (Ferdaus et al., 2018) PAC: A Novel Self-Adaptive Neuro-Fuzzy Controller for Micro Aerial Vehicles
• (Daftry et al., 2016) Robust Monocular Flight in Cluttered Outdoor Environments
• (Kanellakis et al., 2019) Autonomous visual inspection of large-scale infrastructures using aerial robots
• (Mao et al., 2021) Aggressive Visual Perching with Quadrotors on Inclined Surfaces
• (Borowczyk et al., 2016) Autonomous Landing of a Multirotor Micro Air Vehicle on a High Velocity Ground Vehicle
• (Singh et al., 8 Apr 2025) Deep RL-based Autonomous Navigation of Micro Aerial Vehicles (MAVs) in a complex GPS-denied Indoor Environment