Abstract
In this letter, the design, modeling and control of a novel morphing quadrotor are presented. The morphing quadrotor can fly stably and accurately in the air while simultaneously undergoing shape transformation, regardless of the asymmetry of the model. The four arms can rotate around hinges on the main body of the quadrotor to form various topological models. The arms are not in the same plane, so they can overlap with each other. In the extreme case, the width of the morphing quadrotor can be reduced to the diameter of a single rotor to allow the quadrotor to fly through narrow gaps more easily. Reinforcement learning (RL) with an extended-state approach is introduced in this paper to optimize the attitude control law and enable automatic adaptation to model changes. A deterministic policy gradient (DPG) algorithm based on an actor-critic structure with four neural networks in a model-free approach is used to train the controller. Finally, a linear programming method named fast simplex algorithm is presented to solve the control allocation problem of morphing quadrotors in real time with affordable computational cost in this paper. The controller has been tested on our real morphing quadrotor platform and achieves excellent flight performance.
Introduction
Standard quadrotors are inherently rigid, which limits their ability to navigate through narrow vertical or horizontal gaps smaller than their frame. While various morphing drone designs have been proposed to overcome this, most are constrained by symmetrical requirements or low compression ratios. This paper introduces a novel morphing quadrotor capable of shrinking its width to the diameter of a single rotor—the highest compression ratio achieved to date.
The primary challenge in such designs is not merely the mechanical folding mechanism but the complex control physics involved. As the drone transforms, its Center of Gravity (COG) shifts, its Moments of Inertia (MOI) change, and overlapping rotors can lead to aerodynamic interference and lift loss. Traditional controllers, such as standard PID, typically assume a fixed, symmetrical physical model and often fail under these dynamic conditions.
Methodology
The researchers employed a hardware-software co-design to address these challenges.
Mechanical Innovation
The drone features a multi-planar arm design where the four arms are mounted on different horizontal planes. This configuration allows the arms to overlap without physical collision. By alternating rotors between the top and bottom of the arms, the drone can fold into an "Inverted Y" or parallel shape, significantly reducing its footprint.
Extended-State RL Control
To manage the shifting physics, an "extended-state" Reinforcement Learning (RL) controller was developed using a Deterministic Policy Gradient (DPG) algorithm. Unlike conventional controllers, this approach includes the drone's current MOI directly in the input state. This enables the neural network to adapt its control logic instantly based on the drone's physical configuration. To reduce computational overhead, the controller treats the x, y, and z axes as three separate "inverted pendulums," allowing it to run efficiently on onboard hardware.
Control Allocation
Morphing introduces a complex control allocation problem: determining the necessary motor power for specific maneuvers when motors are no longer in standard positions. The authors implemented a "Fast Simplex Algorithm," a linear programming method, to solve this in real-time (under 3ms).
Results and Discussion
The system was validated through both simulation and physical testing using a custom-built platform equipped with a Raspberry Pi 3B and IMU sensors.
Extreme Compression and Stability
The drone successfully demonstrated the ability to transform into highly compressed shapes, including an inverted Y-shape and a parallel configuration where the width is reduced to that of a single rotor.
In simulations, the RL controller significantly outperformed LQR and static PID controllers during rapid MOI changes. The authors also accounted for a 20% loss in lift efficiency during rotor overlap, modeling it as a linear relationship based on rotor proximity.
Real-Time Performance
The RL inference process takes less than 10 microseconds, while the control allocation is executed in under 3 milliseconds. These metrics confirm the practicality of the approach for real-world flight operations.
Conclusion
This work proves that with an intelligent controller, highly asymmetrical drones can achieve excellent flight performance. Future work will focus on integrating visual sensors to enable autonomous gap perception and shape transformation in unknown environments.
Summary updated