Hybrid Neural Network and Conventional Controller Approach for Robust Control of Highly Unstable Systems: Application to Tilt-Rotor Control

Technical Summary: Hybrid Neural Network and Conventional Controller Approach for Robust Control of Highly Unstable Systems

Problem Statement
Conventional multirotor unmanned aerial vehicles (UAVs) are inherently under-actuated, limiting their maneuverability. Tilt-rotor configurations address this by enabling full actuation through thrust-vectoring inputs, allowing independent control over all degrees of freedom (DoF). However, these systems introduce significant dynamic complexity and instability. While deep learning has been integrated into various control strategies (e.g., PID, Sliding Mode Control, MPC), applying direct input-output learning to highly unstable plants remains a challenge. This paper investigates whether neural networks can directly stabilize a fully actuated tilt-rotor system and proposes a hybrid alternative that combines neural network-based system identification with robust conventional control.

Methodology
The research is structured into two distinct phases:

1. Direct Input-Output Learning (Negative Result):
   The authors first attempted to control the tilt-rotor system using a direct mapping approach. In this method, Multilayer Perceptrons (MLP), Long Short-Term Memory (LSTM) networks, and Transformer models were trained to map system states and desired trajectories directly to control inputs (motor speeds and tilt angles). The training data was generated via sinusoidal trajectories and specific maneuvers using a pre-designed Sliding Mode Controller (SMC). The goal was to evaluate if these networks could learn the control policy without explicit dynamic modeling.

2. Neural-Enhanced Sliding Mode Control (Main Contribution):
   Recognizing the failure of the direct approach, the authors proposed a hybrid control architecture. Instead of learning the control policy directly, the neural networks are used to approximate specific components of the plant's dynamic model.

   • Decomposition: The system dynamics are decomposed into input-independent ($a$) and input-dependent ($G$) components based on the equation $\ddot{\bar{x}} = a + Gv$.
   • Learning: Lightweight neural networks (MLP or LSTM) are trained to predict $\hat{a}$ and $\hat{G}$ using a small dataset derived from flight logs or simulations. These networks learn the relationship between system states (attitude, angular velocity) and the dynamic parameters.
   • Control Law: The predicted parameters ($\hat{a}, \hat{G}$) are embedded into a standard Sliding Mode Controller (SMC). The SMC uses these learned approximations to calculate the control law, ensuring robustness against uncertainties.
   • Actuator Allocation: The controller outputs a virtual input vector $v$, which is then mapped to physical actuator commands (BLDC motor speeds and servo tilt angles) using a minimum-norm solution to handle the non-square nature of the control matrix.

Key Results

• Failure of Direct Learning: The direct input-output control strategy failed to stabilize the highly unstable tilt-rotor system. Despite achieving low Mean Absolute Error (MAE) on training data, the networks could not prevent error accumulation in real-time, leading to system divergence. This highlights the fundamental limitation of direct learning for unstable plants without an underlying robust control framework.
• Success of Hybrid Approach: The neural-enhanced SMC successfully stabilized the system.
  • Robustness: The system maintained stability under significant uncertainties, including 5-10% random actuation errors, 5% mass increases, 20% inertia variations, and external wind disturbances.
  • Performance Comparison: Between the two network architectures used for dynamic prediction, the LSTM-based Plant Dynamics Predictor (PDP) outperformed the MLP-based PDP. The LSTM controller achieved a lower Trajectory Tracking Mean Absolute Error (TT-MAE) of 0.21 in challenged conditions compared to 1.05 for the MLP.
  • Efficiency: The LSTM implementation also demonstrated lower runtime (38.5s vs. 41.5s for the baseline test) and required fewer parameters for the prediction of the $a$ component, making it suitable for deployment on low-cost microcontrollers.

Significance and Claims
The paper claims that while direct neural network control is unsuitable for highly unstable tilt-rotor systems, integrating lightweight neural networks into a model-based control framework offers a robust and efficient solution. Key contributions include:

• Demonstrating Limitations: Providing empirical evidence that direct input-output learning fails for highly unstable plants, even with advanced architectures like Transformers.
• Data Efficiency: The proposed method can be trained using flight logs collected from low-performance, attitude-only controllers, as it focuses on learning plant parameters rather than control actions.
• Simulation Utility: The dynamic model learned from real-world data can be utilized for accurate simulations.
• Computational Feasibility: By decomposing the dynamics and using lightweight networks for specific components, the method reduces real-time computational demands, enabling deployment on resource-constrained hardware.

The authors conclude that this hybrid approach provides a generalized, robust control strategy for fully actuated UAVs and plan to validate these findings experimentally on a physical tilt-rotor platform in their laboratory.