• Unified integration of semantic SLAM, coverage path planning, and change detection into a single framework for underwater multi-session perception-aware coverage path planning, enabling autonomous underwater robots to efficiently explore, map, and monitor dynamic environments over multiple sessions. • Integration of a semantic SLAM algorithm for accurate detection and localization of targets. • Evaluation of two distinct coverage policies designed for the discovery and the change detection missions, enabling mission-specific AUV behavior. • Implementation of a semantic-based change detection algorithm to monitor and assess target states across multiple monitoring sessions. • Employment of a probabilistic approach to build the map of the surrounding environment, taking into account both the occupancy probability and the accuracy of identifying an OPI. • Simulation and experimental results to evaluate the consistency of the proposed strategy in terms of target localization, area coverage and change detection, as well as its applicability to real-world scenarios. This paper introduces a novel framework for multi-session, perception-aware coverage path planning integrated with active semantic Simultaneous Localization and Mapping (SLAM) and automatic change detection. The goal is to enhance autonomous robotic exploration in dynamic environments by combining semantic understanding with adaptive path planning for long-term monitoring. The proposed approach consists of three tightly integrated components. First, a semantic-informed inspection planner uses Kernel Density Estimation (KDE) to prioritize exploration of semantically significant regions. Second, an active semantic SLAM module builds a semantic map incrementally, providing real-time feedback to refine the inspection path. Third, a multi-session change detection strategy compares current and previous semantic data to identify and localize environmental changes. Together, these components allow the robot to intelligently adapt its exploration strategy over time, focusing on areas of interest and reacting to environmental dynamics. The framework is validated through both simulation and real-world experiments, demonstrating improved coverage efficiency, mapping accuracy, and change detection robustness compared to traditional methods. While applied to buoy detection, the system is broadly applicable to long-term robotic tasks such as environmental monitoring, infrastructure inspection, mine counter-measure operations, and disaster response, or other scenarios that demand adaptability and semantic awareness in complex, evolving environments.
Bucci et al. (Wed,) studied this question.