SA-MPC: Situational Awareness-Driven Model Predictive Control via Vision-SBES Fusion for Robust Underwater Obstacle Avoidance

Abstract

Given the complexity and uncertainty of underwater environments, achieving situational awareness (SA) and incorporating it into the control loop are essential for ensuring the safety of autonomous underwater vehicles (AUVs). This article presents a novel SA-model predictive control (MPC) framework that integrates vision-based perception and single beam echo sounder (SBES) sensing for robust obstacle avoidance. The proposed system consists of three key modules: 1) a depth fusion module that combines monocular depth estimation with sparse SBES range data to generate high-resolution, metric-consistent depth maps; 2) an SA perception module that infers environmental safety levels through spatiotemporal semantic reasoning, producing pixel-level safety maps; and 3) a learning-based MPC controller that embeds SA-derived spatial risk potentials into optimization process to achieve adaptive decision-making. Both simulation and real-world experiments are conducted to validate the proposed framework. The depth fusion module achieves a mean relative error of 1.65% for the global absolute depth estimation. The SA module effectively identifies dynamic and deformable obstacles, demonstrating improved environmental comprehension under visually degraded conditions. Furthermore, the SA-MPC controller achieves larger obstacle clearance, smoother trajectories, and lower energy consumption compared with traditional A*+MPC approaches. Overall, the proposed SA-MPC framework provides a cost-effective, perception-driven control strategy that enhances the lightweight AUVs operating in complex open-water environments.