Abstract

We introduce OceanGym, the first comprehensive benchmark for ocean underwaterembodied agents, designed to advance AI in one of the most demanding real-worldenvironments. Unlike terrestrial or aerial domains, underwater settings presentextreme perceptual and decision-making challenges, including low visibility,dynamic ocean currents, making effective agent deployment exceptionallydifficult. OceanGym encompasses eight realistic task domains and a unifiedagent framework driven by Multi-modal Large Language Models (MLLMs), whichintegrates perception, memory, and sequential decision-making. Agents arerequired to comprehend optical and sonar data, autonomously explore complexenvironments, and accomplish long-horizon objectives under these harshconditions. Extensive experiments reveal substantial gaps betweenstate-of-the-art MLLM-driven agents and human experts, highlighting thepersistent difficulty of perception, planning, and adaptability in oceanunderwater environments. By providing a high-fidelity, rigorously designedplatform, OceanGym establishes a testbed for developing robust embodied AI andtransferring these capabilities to real-world autonomous ocean underwatervehicles, marking a decisive step toward intelligent agents capable ofoperating in one of Earth's last unexplored frontiers. The code and data areavailable at https://github.com/OceanGPT/OceanGym.