Neural Robust Control on Lie Groups Using Contraction Methods (Extended Version)

A team of researchers from the University of Toronto Institute for Aerospace Studies (UTIAS) has published a novel framework for synthesizing provably robust AI controllers for complex robotic systems. The paper, "Neural Robust Control on Lie Groups Using Contraction Methods," introduces a method to jointly train a neural network-based feedback controller alongside a Robust Control Contraction Metric (RCCM). This dual training ensures the system's behavior respects the underlying geometric structure—a Lie group—which is essential for accurately modeling rotations and orientations in robots like drones and robotic arms.

The core innovation is providing formal stability guarantees for neural controllers on these non-Euclidean spaces. The framework mathematically derives sufficient conditions for the existence of a controller and metric pair that enforces contraction, creating a bounded "tube" around desired trajectories that accounts for external disturbances. In a practical demonstration, the team applied their framework to design a flight controller for a quadrotor. Numerical simulations showed this neural controller outperformed a conventional geometric controller, offering a template for creating more reliable and certifiable AI-driven autonomy in safety-critical applications.