Säker återvändningsplanering för autonoma undervattensfarkoster

Autonomous Underwater Vehicles (AUVs) are often deployed for longduration missions such as seabed mapping, environmental monitoring, or search and rescue. A fundamental requirement in such operations is that the vehicle not only reaches its goal but also guarantees a safe and energy-feasible return to its base. We present a sampling-based motion planning algorithm that extends Rapidly-exploring Random Tree* (RRT*) methods with integrated safety constraints to ensure guaranteed return feasibility throughout the mission. Our approach is motivated by scenarios where autonomous agents, such as underwater vehicles or ground robots, must not only reach a goal but also guarantee a feasible and collision-free return to a base location under motion and energy constraints. Unlike traditional RRT* algorithms that plan unidirectionally to a goal, our planner grows a tree that explicitly checks for the existence of return paths from every sampled node back to the start. These return paths are validated using Dubins vehicle constraints and cached for efficiency. As a result, any node added to the tree is guaranteed to have a safe path backwards. This allows the agent to explore the environment with strong retrieval guarantees without requiring additional post-processing or reactive replanning. In addition to this offline planner, we introduce a lightweight online replanning strategy that handles unexpected disturbances or changes in the environment. When localization error or new obstacles invalidate the planned trajectory, our reconnection algorithm rapidly computes a new feasible path back to both the origin and the goal by reattaching the disturbed state to the existing tree. This enables robust recovery and mission continuity in dynamic settings. The algorithm is evaluated in randomized simulation environments across various challenging scenarios, including disturbances in state estimation and the sudden appearance of new obstacles. We analyze performance using metrics such as path optimality, curvature percentage, reconnection success rate, and computation time. Experimental results demonstrate high success rates and rapid replanning capabilities, with asymptotically improving path quality. Our approach provides a principled, efficient framework for safetycritical motion planning in dynamic and constrained environments.