Introduction
Quadrotors are inherently under-determined systems, typically requiring four functional rotors to maintain six-degree-of-freedom control. The sudden failure of a single rotor usually results in a catastrophic loss of stability, leading to a crash. To mitigate this, researchers have historically developed Fault-Tolerant Control (FTC) systems, categorized into two main types: Active (AFTC) and Passive (PFTC).
Active systems rely on a Fault Detection and Diagnosis (FDD) module to identify a failure and trigger a specialized recovery controller. However, the latency inherent in FDD—often exceeding 0.1 seconds—can be fatal in high-speed scenarios. Conversely, Passive systems treat faults as generalized disturbances. While they avoid switching delays, their "one-size-fits-all" design often results in degraded performance during both normal and faulty flight.
This paper introduces a hybrid approach: a "Selector-Controller" neural network architecture. This system achieves the high-fidelity performance of an active expert while maintaining the seamless, delay-free robustness of a passive controller.
Methodology
Selector-Controller Architecture
The proposed architecture moves away from a single monolithic policy. Instead, it utilizes a modular structure consisting of:
- Four Controller Networks: Each is a Multi-Layer Perceptron (MLP) optimized for a specific failure mode (e.g., Rotor 1 failure, Rotor 2 failure, etc.).
- A Selector Network: This module processes the drone's instantaneous state—including linear acceleration and angular velocity—to output a set of weights.
The final control command is a weighted sum of the outputs from the four controllers. This design effectively mitigates "gradient conflict," a common training issue where the optimization for one failure mode inadvertently degrades the performance of another.
Hybrid Training Framework
To train this architecture, the authors employed a multi-faceted objective function:
- Reinforcement Learning (PPO): Proximal Policy Optimization is used to maximize a reward function based on position tracking and stability.
- Behavior Cloning (BC): The network is trained to mimic an "AFTC expert"—a controller with perfect ground-truth knowledge of the fault—allowing the student network to reach expert-level performance.
- Supervised Learning: The selector network is explicitly supervised to correctly identify the failing rotor based on the mismatch between commanded thrust and observed IMU data.
A critical aspect of the control strategy is the "Primary Axis" approach. Upon failure, the system acknowledges that yaw control is physically impossible. It instead prioritizes the "primary axis," tilting the spinning airframe to ensure the remaining thrust vectors maintain altitude and lateral position.
Experimental Results
Simulation and Real-World Validation
The system was rigorously tested using the Flightmare simulator, involving over 1.2 billion training steps and 10,000 randomized failure scenarios. Physical validation was conducted on a custom Q250 quadrotor equipped with a Pixhawk 6C mini and an Orange Pi for onboard computation.
Performance Metrics
The Selector-Controller architecture demonstrated significant improvements over existing benchmarks:
- Precision: In real-world flight tests, the proposed method maintained a Mean Position Error of only 0.100m, whereas standard PFTC methods exceeded 0.5m.
- Latency: By eliminating the explicit FDD switching step, the system reacts to failures in milliseconds, avoiding the "switching shock" or jerks common in AFTC systems.
- Efficiency: The policy is computationally lightweight, requiring only 0.8 ms for inference on an STM32H7 flight controller, making it highly practical for real-time deployment on standard drone hardware.
Conclusion
The Selector-Controller architecture provides a robust solution for quadrotor rotor failure by blending the strengths of active and passive control. By leveraging hybrid training and a modular network structure, the system offers high-precision recovery without the delays of traditional fault detection. Future research will focus on transitioning from external motion capture systems to onboard sensing (such as VIO) and testing the system's limits during high-speed, aggressive maneuvers in outdoor environments.
Summary updated