dual-pathway-robust-learning-control

name	dual-pathway-robust-learning-control
description	Dual-pathway architecture for provably robust learning-based control combining neural network feedforward estimation with conventional feedback correction. Use when designing learning-enabled control systems requiring both performance and robustness guarantees, particularly for robotics, aerospace, quantum control, or any system with out-of-distribution deployment concerns.
metadata	{"arxiv_id":"2607.06535","published":"2026-07-07","authors":"Fan Zhang, Richie Suganda, Jinfeng Chen, Wenhua Liu, Hantao Fu, Bin Hu, Qin Lin","tags":["robust-control","learning-based-control","neural-networks","Lyapunov","ESO","systems-engineering"]}

Dual-Pathway Robust Learning Control

Core Concept

Learning-based control methods often over-rely on learned models, causing failure when deployed out-of-distribution. The Neural-ESO (Neural Extended State Observer) framework solves this with a dual-pathway architecture:

1. Predictive pathway: Neural network provides feedforward disturbance estimate to accelerate convergence
2. Corrective pathway: Conventional ESO compensates prediction errors and prevents over-reliance on neural component

Key theorem: Enforcing a Lipschitz bound on the learning component guarantees uniform ultimate boundedness of the closed-loop error dynamics (via Lyapunov theory + small-gain analysis).

Mathematical Framework

Given system: ẋ = f(x) + g(x)u + d(t) where d(t) is unknown disturbance

Neural-ESO structure:

û = u_nominal - NN(x) - ESO_residual

• NN(x): Neural network feedforward disturbance estimate