Learning-Based Passive Fault-Tolerant Control of a Quadrotor with Rotor Failure

TL;DR

This work addresses quadrotor fault tolerance under arbitrary single-rotor failures without relying on fault information or controller switching. It introduces a learning-based passive fault-tolerant control (PFTC) framework with a Selector-Controller network that jointly infers faults and computes fault-adaptive commands by unifying fault detection and control within a single policy. The approach blends RL, BC, and supervised learning, and demonstrates faster fault response and stronger position tracking than state-of-the-art PFTC methods, while achieving performance comparable to ideal AFTC that uses fault information. Extensive simulations and real-world experiments validate the method's robustness across severity levels and its practical viability on real hardware with real-time constraints. The solution offers a scalable path toward resilient quadrotor operation without reliance on accurate fault diagnostics or explicit controller switching.

Abstract

This paper proposes a learning-based passive fault-tolerant control (PFTC) method for quadrotor capable of handling arbitrary single-rotor failures, including conditions ranging from fault-free to complete rotor failure, without requiring any rotor fault information or controller switching. Unlike existing methods that treat rotor faults as disturbances and rely on a single controller for multiple fault scenarios, our approach introduces a novel Selector-Controller network structure. This architecture integrates fault detection module and the controller into a unified policy network, effectively combining the adaptability to multiple fault scenarios of PFTC with the superior control performance of active fault-tolerant control (AFTC). To optimize performance, the policy network is trained using a hybrid framework that synergizes reinforcement learning (RL), behavior cloning (BC), and supervised learning with fault information. Extensive simulations and real-world experiments validate the proposed method, demonstrating significant improvements in fault response speed and position tracking performance compared to state-of-the-art PFTC and AFTC approaches.

Figures (10)

• Figure 1: An overview of our policy training framework. The dashed lines represent updates to the policy parameters, while the solid lines indicate the flow of relevant variables. We train the policy using a combination of Reinforcement Learning (RL), Behavior Cloning (BC), and supervised learning with fault information.
• Figure 2: Definition of quadrotor body frame and rotor indices.
• Figure 3: Control structure of high-level controller and low-level controller.
• Figure 4: The Selector-Controller network structure consists of identical MLP architectures for both the selector and controller networks. The final control output is computed as a weighted sum of the outputs from all controller networks, where the weights are determined by the selector network.
• Figure 5: Time history of the quadrotor states controlled by various methods in the simulations. "INDI" and "NMPC" methods are the AFTC methods. "ideal" represents a controller switch was performed simultaneously with the fault trigger and "FDD" represents that an FDD module was conducted to detect faults and switch controllers as needed. "Uniform PFTC" and "Ours" methods are the PFTC methods, which do not require a controller switch or prior knowledge of fault information.