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

Researchers demonstrate that direct neural network approaches fail for controlling highly unstable tilt-rotor systems, but propose a hybrid solution combining sliding mode control with neural networks to predict system dynamics. The LSTM-based implementation outperforms traditional methods while reducing computational overhead, advancing autonomous aerial vehicle control capabilities.

This research addresses a fundamental challenge in control systems: stabilizing inherently unstable plants using machine learning. The paper's two-part structure—first documenting why direct neural network approaches fail, then presenting a successful hybrid alternative—provides valuable engineering insights often absent from publications that only report successes. The negative result carries significant weight for the control systems and robotics communities, establishing that MLPs, LSTMs, and transformers cannot simply be trained as black-box controllers for unstable aerospace systems. The proposed solution decomposes the problem intelligently: rather than attempting end-to-end learning, it uses neural networks only to model the input-independent dynamics while conventional sliding mode control handles stabilization. This hybrid approach leverages the strengths of both paradigms—classical control guarantees stability while neural networks improve precision through learned dynamics from real flight data. The use of LSTM networks specifically demonstrates that recurrent architectures better capture temporal dependencies in system behavior compared to feedforward alternatives, achieving both superior performance and lower computational cost. For the autonomous vehicle industry, this work has practical implications: it shows how to incorporate machine learning into safety-critical systems without abandoning mathematical guarantees. The method's ability to learn from suboptimal controller logs rather than requiring expert demonstrations reduces training data requirements, accelerating deployment timelines. As tilt-rotor and vectored-thrust systems become more prevalent in commercial and military applications, control methods that balance robustness with learning capability directly impact system reliability and operational flexibility.

• →Direct neural network control fails for unstable systems, but hybrid approaches combining classical control with learned dynamics succeed.
• →LSTM-based dynamic predictors outperform MLP alternatives while demanding less computational resources for real-time control.
• →The method trains on suboptimal flight data rather than expert demonstrations, reducing expensive data collection requirements.
• →Decomposing control into input-independent and input-dependent components allows neural networks to focus on modeling rather than stabilization.
• →This hybrid framework preserves mathematical stability guarantees while incorporating the adaptability benefits of machine learning.