A Differentiable QP-based Learning Framework for Safety-Critical Control of Fully Actuated AUVs

This paper tackles the safety-critical control problem of the fully actuated autonomous underwater vehicle (AUV) in the presence of unknown disturbances. Specifically, a differentiable quadratic program (QP)-based learning framework that integrates fully connected layers and high order control barrier function (HOCBF) is proposed to avoid the obstacle. The safety-critical loss is designed by considering the quantified safety violations to resist unknown disturbances during the training process. In the control loop, a first-order high-gain disturbance observer (DOB) is employed to estimate the unknown dynamics associated with the forward invariance condition. This approach effectively diminishes the system’s conservativeness by compensating for the bounded estimation error of the DOB. The simulation results demonstrate that the proposed framework enables the AUV to satisfy safety requirements in the presence of unknown disturbances, while reducing the dependence of the system performance on artificial parameters by adaptively regulating the class $\mathcal{K}$ functions through neural network outputs.