Underwater Robotic Systems with Real-Time Acoustic AI Mapping

Abstract

Underwater robotic exploration has been revolutionized by advances in acoustic sensing and artificial intelligence (AI), enabling autonomous underwater vehicles (AUVs) to navigate and map complex seabed environments without reliance on optical systems. Traditional optical approaches fail under turbid conditions and at depth, where light attenuation is severe, whereas acoustic sensors—such as multi‑beam and side‑scan sonar—provide robust, range‐based measurements regardless of visibility. However, raw sonar data are plagued by speckle noise, multipath reflections, and environmental interference, which complicate feature extraction and mapping. To address these challenges, we present Underwater Robotic Systems with Real‑Time Acoustic AI Mapping, a comprehensive framework that integrates a synchronized multi‑modal acoustic sensor suite with onboard deep learning models for semantic feature extraction and simultaneous localization and mapping (SLAM). Our system hardware combines a multi‑beam sonar for broad bathymetric coverage, side‑scan sonar for high-resolution imagery, an inertial measurement unit (IMU) for attitude estimation, and a depth sensor for accurate pressure-based altitude readings. Sensor streams are synchronized using Precision Time Protocol (PTP) to ensure consistent timestamp alignment. At the core of our software pipeline lies a convolutional neural network (CNN) based on the U‑Net architecture, trained to segment seabed structures—such as ridges, pipelines, and wreckage—in acoustic intensity images. The network was trained on a curated dataset of over 5,000 labeled sonar returns, augmented with simulated noise profiles replicating real‑world interference patterns.