PolarNet guarantees single minimum for provably stable neural control

A team of researchers from Yuan Zhong and colleagues has introduced PolarNet, a neural network architecture designed to overcome a fundamental limitation in learning provably stable controllers. Existing neural Lyapunov control methods often suffer from training-time behaviors that hinder the synthesis of a stable controller, particularly under specific conditions. The root cause, the team identified, is the lack of architectural guarantees on the learned Lyapunov function—standard networks can produce multiple spurious critical points that mislead optimization. PolarNet structurally guarantees a single critical point (minimum), a key requirement for Lyapunov functions to certify stability. The authors provide theoretical proofs of properness and universality, showing that PolarNet can represent any valid Lyapunov function while maintaining this property.

When used as a drop-in replacement in existing neural Lyapunov control methods, PolarNet effectively circumvents training failures caused by these spurious minima. In numerical experiments across various dynamical systems, PolarNet consistently maintained a single critical point during training, while standard architectures produced multiple minima that led to optimization failures. The approach does not require changes to the training objectives or algorithms—teams can simply swap the network architecture. This makes PolarNet a practical tool for engineers building safe, provably stable controllers for robotics, drones, and autonomous vehicles, where stability guarantees are critical. The code is publicly available, allowing researchers to test the method on their own control problems.