README.md

metadata

language:
  - en
license: mit
task_categories:
  - robotics
tags:
  - agent
  - robotics
  - benchmark
  - environment
  - underwater
  - multi-modal
  - mllm
  - large-language-models

🌊 OceanGym 🦾

A Benchmark Environment for Underwater Embodied Agents

🌐 Home Page 📄 Paper 💻 Code 🤗 Hugging Face ☁️ Google Drive ☁️ Baidu Drive

OceanGym is a high-fidelity embodied underwater environment that simulates a realistic ocean setting with diverse scenes. As illustrated in figure, OceanGym establishes a robust benchmark for evaluating autonomous agents through a series of challenging tasks, encompassing various perception analyses and decision-making navigation. The platform facilitates these evaluations by supporting multi-modal perception and providing action spaces for continuous control.

We introduce OceanGym, the first comprehensive benchmark for ocean underwater embodied agents, designed to advance AI in one of the most demanding real-world environments. Unlike terrestrial or aerial domains, underwater settings present extreme perceptual and decision-making challenges, including low visibility, dynamic ocean currents, making effective agent deployment exceptionally difficult. OceanGym encompasses eight realistic task domains and a unified agent framework driven by Multi-modal Large Language Models (MLLMs), which integrates perception, memory, and sequential decision-making. Agents are required to comprehend optical and sonar data, autonomously explore complex environments, and accomplish long-horizon objectives under these harsh conditions. Extensive experiments reveal substantial gaps between state-of-the-art MLLM-driven agents and human experts, highlighting the persistent difficulty of perception, planning, and adaptability in ocean underwater environments. By providing a high-fidelity, rigorously designed platform, OceanGym establishes a testbed for developing robust embodied AI and transferring these capabilities to real-world autonomous ocean underwater vehicles, marking a decisive step toward intelligent agents capable of operating in one of Earth's last unexplored frontiers. The code and data are available at this https URL .

💐 Acknowledgement

OceanGym environment is based on Unreal Engine (UE) 5.3.

Partial functions of OceanGym is developed on HoloOcean.

Thanks for their great contributions!

🔔 News

• 09-2025, we launched the OceanGym project.
• 08-2025, we finshed the OceanGym environment.

---

Contents:

• 💐 Acknowledgement
• 🔔 News
• 📺 Quick Start
  • Decision Task
  • Perception Task
• ⚙️ Set up Environment
  • Clone HoloOcean
  • Packaged Installation
  • Add World Files
  • Open the World
• 🧠 Decision Task
  • Target Object Locations
  • Evaluation Criteria
• 👀 Perception Task
  • Using the Bench to Eval
    • Import Data
    • Set your Model Parameters
    • Simple Multi-views
    • Multi-views with Sonar
    • Multi-views add Sonar Examples
  • Collecting Image Data
    • Modify Configuration File
    • Collect Camera Images Only
    • Collect Camera and Sonar Images
• ⏱️ Results
  • Decision Task
  • Perception Task
• 📚 Datasets
• 🚩 Citation

📺 Quick Start

Install the experimental code environment using pip:

pip install -r requirements.txt

Decision Task

Only the environment is ready! Build the environment based on here.

Step 1: Run a Task Script

For example, to run task 4:

python decision\tasks\task4.py

Follow the keyboard instructions or switch to LLM mode for automatic decision-making.

Step 2: Keyboard Control Guide

Key	Action
W	Move Forward
S	Move Backward
A	Move Left
D	Move Right
J	Turn Left
L	Turn Right
I	Move Up
K	Move Down
M	Switch to LLM Mode
Q	Exit

You can use WASD for movement, J/L for turning, I/K for up/down. Press M to switch to large language model mode (may cause temporary lag). Press Q to exit.

Step 3: View Results

Logs and memory files are automatically saved in the log/ and memory/ directories.

Step 4: Evaluate the results

Place the generated memory and important_memory files into the corresponding point folders. Then, set the evaluation paths in the evaluate.py file.

We provide 6 experimental evaluation paths. In evaluate.py, you can configure them as follows:

eval_roots = [
    os.path.join(eval_root, "main", "gpt4omini"),
    os.path.join(eval_root, "main", "gemini"),
    os.path.join(eval_root, "main", "qwen"),
    os.path.join(eval_root, "migration", "gpt4o"),
    os.path.join(eval_root, "migration", "qwen"),
    os.path.join(eval_root, "scale", "qwen"),
]

To run the evaluation:

python decision\utils\evaluate.py

The generated results will be saved under the \eval\decision folder.

Perception Task

Step 1: Prepare the dataset

After downloading from Hugging Face, and put it into the data/perception folder.

Step 2: Select model parameters

parameter	function
model_template	The large language model message queue template you selected.
model_name_or_path	If it is an API model, it is the model name; if it is a local model, it is the path.
api_key	If it is an API model, enter your key.
base_url	If it is an API model, enter its baseful URL.

Now we only support OpenAI, Google Gemma, Qwen and OpenBMB.

MODELS_TEMPLATE="Yours"
MODEL_NAME_OR_PATH="Yours"
API_KEY="Yours"
BASE_URL="Yours"

Step 3: Run the experiments

parameter	function
exp_name	Customize the name of the experiment to save the results.
exp_idx	Select the experiment number, or enter "all" to select all.
exp_json	JSON file containing the experiment label data.
images_dir	The folder where the experimental image data is stored.

For the experimental types, We designed (1) multi-view perception task and (2) context-based perception task.

For the lighting conditions, We designed (1) high illumination and (2) low illumination.

For the auxiliary sonar, We designed (1) without sonar image (2) zero-shot sonar image and (3) sonar image with few sonar example.

Such as this command is used to evaluate the multi-view perception task under high illumination:

python perception/eval/mv.py \
    --exp_name Result_MV_highLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLight.json" \
    --images_dir "/data/perception/highLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

For more patterns about perception tasks, please read this part carefully.

⚙️ Set up Environment

This project is based on the HoloOcean environment. 💐

We have placed a simplified version here. If you encounter any detailed issues, please refer to the original installation document.

Clone HoloOcean

Make sure your GitHub account is linked to an Epic Games account, please Follow the steps here and remember to accept the email invitation from Epic Games.

After that clone HoloOcean:

git clone git@github.com:byu-holoocean/HoloOcean.git holoocean

Packaged Installation

1. Additional Requirements

For the build-essential package for Linux, you can run the following console command:

sudo apt install build-essential

2. Python Library

From the cloned repository, install the Python package by doing the following:

cd holoocean/client
pip install .

3. Worlds Packages

To install the most recent version of the Ocean worlds package, open a Python shell by typing the following and hit enter:

python

Install the package by running the following Python commands:

import holoocean
holoocean.install("Ocean")

To do these steps in a single console command, use:

python -c "import holoocean; holoocean.install('Ocean')"

Add World Files

Place the JSON config file from asset/decision/map_config or asset\perception\map_config into some place like:

(Windows)

C:\Users\Windows\AppData\Local\holoocean\2.0.0\worlds\Ocean

Open the World

1. If you're use it in first time, you have to compile it

1-1. find the Holodeck.uproject in engine folder

1-2. Right-click and select:Generate Visual Studio project files

1-3. If the version is not 5.3.2,please choose the Switch Unreal Engine Version

1-4. Then open the project

2. Then find the HAIDI map in demo directory

3. Run the project

🧠 Decision Task

All commands are applicable to Windows only, because it requires full support from the UE5 Engine.

The decision experiment can be run with reference to the Quick Start.

Target Object Locations

We have provided eight tasks. For specific task descriptions, please refer to the paper.

The following are the coordinates for each target object in the environment (in meters):

• MINING ROBOT: (-71, 149, -61), (325, -47, -83)
• OIL PIPELINE: (345, -165, -32), (539, -233, -42), (207, -30, -66)
• OIL DRUM: (447, -203, -98)
• SUNKEN SHIP: (429, -151, -69), (78, -11, -47)
• ELECTRICAL BOX: (168, 168, -65)
• WIND POWER STATION: (207, -30, -66)
• AIRCRAFT WRECKAGE: (40, -9, -54), (296, 78, -70), (292, -186, -67)
• H-MARKED LANDING PLATFORM: (267, 33, -80)

---

Evaluation Criteria

1. If the target is not found, use the final stopping position for evaluation.
2. If the target is found, use the closest distance to any target point.
3. For found targets:
   • Minimum distance ≤ 30: full score
   • 30 < distance < 100: score decreases proportionally
   • Distance ≥ 100: score is 0
4. Score composition:
   • One point: 100
   • Two points: 60 / 40
   • Three points: 60 / 20 / 20

👀 Perception Task

Using the Bench to Eval

All commands are applicable to Linux, so if you using Windows, you need to change the corresponding path representation (especially the slash).

Now we only support OpenAI, Google Gemma, Qwen and OpenBMB. If you need to customize the model, please contact the author.

Import Data

First, you need download our data from Hugging Face.

And then create a new data folder in the project root directory:

mkdir -p data/perception

Finally, put the downloaded data into the corresponding folder.

Set your Model Parameters

Just open a terminal in the root directory and set it directly.

parameter	function
model_template	The large language model message queue template you selected.
model_name_or_path	If it is an API model, it is the model name; if it is a local model, it is the path.
api_key	If it is an API model, enter your key.
base_url	If it is an API model, enter its baseful URL.

MODELS_TEMPLATE="Yours"
MODEL_NAME_OR_PATH="Yours"
API_KEY="Yours"
BASE_URL="Yours"

Simple Multi-views

All of these scripts evaluate the perception task, and the parameters are as follows:

parameter	function
exp_name	Customize the name of the experiment to save the results.
exp_idx	Select the experiment number, or enter "all" to select all.
exp_json	JSON file containing the experiment label data.
images_dir	The folder where the experimental image data is stored.

This command is used to evaluate the multi-view perception task under high illumination:

python perception/eval/mv.py \
    --exp_name Result_MV_highLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLight.json" \
    --images_dir "/data/perception/highLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under high illumination:

python perception/eval/mv.py \
    --exp_name Result_MV_highLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLightContext.json" \
    --images_dir "/data/perception/highLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the multi-view perception task under low illumination:

python perception/eval/mv.py \
    --exp_name Result_MV_lowLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLight.json" \
    --images_dir "/data/perception/lowLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under low illumination:

python perception/eval/mv.py \
    --exp_name Result_MV_lowLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLightContext.json" \
    --images_dir "/data/perception/lowLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

Multi-views with Sonar

This command is used to evaluate the multi-view perception task under high illumination with sonar image:

python perception/eval/mvs.py \
    --exp_name Result_MVwS_highLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLight.json" \
    --images_dir "/data/perception/highLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under high illumination with sonar image:

python perception/eval/mvs.py \
    --exp_name Result_MVwS_highLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLightContext.json" \
    --images_dir "/data/perception/highLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the multi-view perception task under low illumination with sonar image:

python perception/eval/mvs.py \
    --exp_name Result_MVwS_lowLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLight.json" \
    --images_dir "/data/perception/lowLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under low illumination with sonar image:

python perception/eval/mvs.py \
    --exp_name Result_MVwS_lowLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLightContext.json" \
    --images_dir "/data/perception/lowLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

Multi-views add Sonar Examples

This command is used to evaluate the multi-view perception task under high illumination with sona image examples:

python perception/eval/mvsex.py \
    --exp_name Result_MVwSss_highLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLight.json" \
    --images_dir "/data/perception/highLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under high illumination with sona image examples:

python perception/eval/mvsex.py \
    --exp_name Result_MVwSss_highLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/highLightContext.json" \
    --images_dir "/data/perception/highLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the multi-view perception task under low illumination with sona image examples:

python perception/eval/mvsex.py \
    --exp_name Result_MVwSss_lowLight_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLight.json" \
    --images_dir "/data/perception/lowLight" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

This command is used to evaluate the context-based perception task under low illumination with sona image examples:

python perception/eval/mvsex.py \
    --exp_name Result_MVwSss_lowLightContext_00 \
    --exp_idx "all" \
    --exp_json "/data/perception/lowLightContext.json" \
    --images_dir "/data/perception/lowLightContext" \
    --model_template $MODELS_TEMPLATE \
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --api_key $API_KEY \
    --base_url $BASE_URL

Collecting Image Data

This part is optional. Only use when you need to collect pictures by yourself.

Modify Configuration File

The sample configuration files can be found in asset/perception/map_config. You need to copy this and paste it into your HoloOcean project's configuration.

Collect Camera Images Only

This command is used to collect camera images only, and the parameters are as follows:

parameter	function
scenario	The name of the json configuration file you want to replace.
task_name	Customize the name of the experiment to save the results.
rgbcamera	The camera directions you can choose. If select all, enter "all".

python perception/task/init_map.py \
    --scenario without_sonar \
    --task_name "Exp_Camera_Only" \
    --rgbcamera "all"

Collect Camera and Sonar Images

This command is used to collect both camera images and sonar images at same time:

python perception/task/init_map_with_sonar.py \
    --scenario with_sonar \
    --task_name "Exp_Add_Sonar" \
    --rgbcamera "FrontCamera"

⏱️ Results

We provide the trajectory data of OceanGym’s various task evaluations at the next section, enabling readers to analyze and reproduce the results.

Decision Task

• This table is the performance in decision tasks requiring autonomous completion by MLLM-driven agents.

Perception Task

• This table is the performance of perception tasks across different models and conditions.
• Values represent accuracy percentages.
• Adding sonar means using both RGB and sonar images.

📚 Datasets

The link to the dataset is as follows
☁️ Google Drive

• Decision Task

decision_dataset
├── main
│ ├── gpt4omini
│ │ ├── task1
│ │ │ ├── point1
│ │ │ │ ├── llm_output_...log
│ │ │ │ ├── memory_...json
│ │ │ │ └── important_memory_...json
│ │ │ └── ... (other data points like point2, point3...)
│ │ └── ... (other tasks like task2, task3...)
│ ├── gemini
│ │ └── ... (structure is the same as gpt4omini)
│ └── qwen
│ └── ... (structure is the same as gpt4omini)
│
├── migration
│ ├── gpt4o
│ │ └── ... (structure is the same as above)
│ └── qwen
│ └── ... (structure is the same as above)
│
└── scale
  ├── qwen
  └── gpt4omini

• Perception Task

perception_dataset
├── data
│ ├── highLight
│ ├── highLightContext
│ ├── lowLight
│ ├── lowLightContext
│
└── result

🚩 Citation

If this OceanGym paper or benchmark is helpful, please kindly cite as this:

@misc{xue2025oceangymbenchmarkenvironmentunderwater,
      title={OceanGym: A Benchmark Environment for Underwater Embodied Agents}, 
      author={Yida Xue and Mingjun Mao and Xiangyuan Ru and Yuqi Zhu and Baochang Ren and Shuofei Qiao and Mengru Wang and Shumin Deng and Xinyu An and Ningyu Zhang and Ying Chen and Huajun Chen},
      year={2025},
      eprint={2509.26536},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.26536}, 
}

General HoloOcean use:

@inproceedings{Potokar22icra,
   author = {E. Potokar and S. Ashford and M. Kaess and J. Mangelson},
   title = {Holo{O}cean: An Underwater Robotics Simulator},
   booktitle = {Proc. IEEE Intl. Conf. on Robotics and Automation, ICRA},
   address = {Philadelphia, PA, USA},
   month = may,
   year = {2022}
}

Simulation of Sonar (Imaging, Profiling, Sidescan) sensors:

@inproceedings{Potokar22iros,
   author = {E. Potokar and K. Lay and K. Norman and D. Benham and T. Neilsen and M. Kaess and J. Mangelson},
   title = {Holo{O}cean: Realistic Sonar Simulation},
   booktitle = {Proc. IEEE/RSJ Intl. Conf. Intelligent Robots and Systems, IROS},
   address = {Kyoto, Japan},
   month = {Oct},
   year = {2022}
}

💐 Thanks again!