OceanPlan is an AI system of piloting AUVs in natural language, which is developed by GTSR Lab of Georgia Tech. Specifically, we establish a hierarchical LLM-task-motion planning and replanning framework to execute robotic control of AUVs in a comprehensive ocean simulator HoloEco.

Autonomous Underwater Vehicles (AUVs) have been extensively studied in the field of marine robotics. In light of the rapid emergence of Large Language Models (LLMs), we aim to lay the foundation for intelligent robotic systems by enabling seamless interaction between AUVs and humans. To achieve such a goal, we must tackle domain-specific difficulties in marine robotics, including difficult underwater localization, uncertainty in ocean environments, and abnormal AUV behaviours. We design a language-guided task and motion planning framework to systematically translate an abstracted human command into primitive AUV control in the ocean. This planning hierarchy endows representations of the world with incremental resolutions, thus efficiently grounding the abstracted command into tangible AUV motions. We also incorporate a holistic replanner to provide real-time feedback with all planners for robust AUV operation. To comprehensively evaluate our framework, we build a photo-realistic marine simulator highlighting a distilled AUV agent. Experiments validate that the proposed framework delivers successful AUV performance of long-duration missions through generalized piloting in natural language.