Neural-Fly enables rapid learning for agile flight in strong winds.

Area of Science:

• Robotics
• Aerospace Engineering
• Machine Learning

Background:

• Safe and precise flight control for Uninhabited Aerial Vehicles (UAVs) is crucial for their widespread adoption.
• Traditional control methods struggle with the complex relationship between dynamic wind conditions and aircraft maneuverability.
• Understanding aerodynamic responses to varied wind is key for designing robust UAV controllers.

Purpose of the Study:

• To develop a learning-based control system, Neural-Fly, for precise UAV flight maneuvers in dynamic high-speed winds.
• To enable rapid online adaptation of UAV controllers by leveraging deep learning and pretrained representations.
• To address the limitations of traditional control design methods in complex wind environments.

Main Methods:

• Implemented a deep learning approach, Neural-Fly, utilizing domain adversarially invariant meta-learning (DAIML).
• Learned a shared aerodynamic representation across different wind conditions, identifying wind-specific variations in a low-dimensional space.
• Employed a composite adaptation law to update linear coefficients for basis element mixing, using only 12 minutes of flight data.

Main Results:

• Neural-Fly demonstrated precise flight control with significantly reduced tracking error compared to state-of-the-art controllers in high-speed winds (up to 43.6 km/h).
• The system achieved exponential stability, providing robustness guarantees for UAV control.
• The control design successfully extrapolated to unseen wind conditions and transferred effectively across different drone platforms.

Conclusions:

• Neural-Fly offers a robust and adaptive solution for UAV flight control in challenging wind conditions.
• The learning-based approach significantly outperforms traditional methods in terms of precision and adaptability.
• The system's ability to generalize and transfer across platforms highlights its potential for real-world UAV applications.