Hypernetwork-Conditioned Reinforcement Learning for Robust Control of Fixed-Wing Aircraft under Actuator Failures

A research team led by Dennis Marquis and Mazen Farhood has published a novel AI control system for unmanned aircraft. Their paper, "Hypernetwork-Conditioned Reinforcement Learning for Robust Control of Fixed-Wing Aircraft under Actuator Failures," details a reinforcement learning (RL) path-following controller specifically designed for small uncrewed aircraft systems (sUAS). The core innovation is the use of a hypernetwork—a neural network that generates the weights for another network—to condition the control policy on a parameterization of potential actuator faults. This allows a single trained model to adapt its behavior dynamically when components fail.

The controller was trained using Proximal Policy Optimization (PPO) and employs parameter-efficient adaptation techniques like Feature-wise Linear Modulation (FiLM) and Low-Rank Adaptation (LoRA) to keep the model lean. Crucially, the team validated the system in high-fidelity simulations using a realistic six-degree-of-freedom aircraft model. The results showed that the hypernetwork-conditioned policies not only improved robustness compared to standard multilayer perceptron (MLP) policies but also demonstrated effective generalization. The AI was able to handle complex, time-varying actuator failure modes that it had never encountered during its initial training phase, a significant step toward reliable autonomous systems in unpredictable environments.