USIM: Underwater Simulation Dataset for Vision-Language-Action Models

TL;DR

USIM is a large-scale underwater robot manipulation and navigation dataset collected in the Stonefish physics simulator. It contains 2,275 episodes (1,750 train + 525 test) across 20 tasks in 9 underwater scenarios, formatted in LeRobot v2.1 format with dual-camera video recordings.

Dataset Description

USIM is introduced in the paper "USIM and U0: A Vision-Language-Action Dataset and Model for General Underwater Robots". It is designed to train and evaluate Vision-Language-Action (VLA) models for autonomous underwater robots operating in diverse subsea environments.

Key Features

• Diverse underwater scenarios: shallow ocean, underwater factory, industrial pool, subsea pipeline, shipwreck sites, lake environments, and open sea
• Dual-camera observation: ego (front-facing) and wrist (end-effector) camera views at 240×320 resolution
• Rich proprioceptive state: 29-dimensional state vector including joint positions, thruster PWM, velocities, IMU data, DVL, and pressure readings
• 20 tasks spanning grasping, navigation, tracking, and transporting

Robot Platform

The robot used is a BlueROV2 underwater vehicle equipped with a 4-DOF robotic arm and a scaled-down Robotiq gripper, simulated in the Stonefish physics engine.

Dataset Structure

This repository contains two independent LeRobot v2.1 datasets:

usim/
├── train/                    # Training split (1,750 episodes)
│   ├── meta/
│   │   ├── info.json
│   │   ├── tasks.jsonl
│   │   ├── episodes.jsonl
│   │   ├── episodes_stats.jsonl
│   │   └── modality.json
│   ├── data/
│   │   ├── chunk-000/
│   │   └── chunk-001/
│   └── videos/
│       ├── chunk-000/
│       │   ├── observation.images.ego/
│       │   └── observation.images.wrist/
│       └── chunk-001/
│           ├── observation.images.ego/
│           └── observation.images.wrist/
├── test/                     # Test split (525 episodes)
│   ├── meta/
│   ├── data/
│   │   └── chunk-000/
│   └── videos/
│       └── chunk-000/
│           ├── observation.images.ego/
│           └── observation.images.wrist/
└── README.md

Supported Tasks

The dataset covers 20 tasks and 9 language instructions grouped into 4 categories:

Grasping

Task Code	Instruction	Scenario
pick_pipe0_shallow	Pick up the pipe	Shallow ocean
pick_pipe1_shallow	Pick up the pipe	Shallow ocean
pick_pipe0_factory	Pick up the pipe	Underwater factory
pick_pipe1_factory	Pick up the pipe	Underwater factory
pick_red_shallow	Pick up the red cylinder	Shallow ocean
pick_redx_shallow	Pick up the red cylinder	Shallow ocean (multi-blue distractors)
pick_red_factory	Pick up the red cylinder	Underwater factory
pick_redx_factory	Pick up the red cylinder	Underwater factory (multi-blue distractors)
pick_blue_shallow	Pick up the blue cylinder	Shallow ocean
pick_bluex_shallow	Pick up the blue cylinder	Shallow ocean (multi-red distractors)
pick_blue_factory	Pick up the blue cylinder	Underwater factory
pick_bluex_factory	Pick up the blue cylinder	Underwater factory (multi-red distractors)

Navigation

Task Code	Instruction	Scenario
goto_charge_station	Go to the charge station	Lake with equipment
goto_water_tower	Go to the water tower	Lake with rocks
scan_ship_modern	Scan the ship	Modern shipwreck
scan_ship_ancient	Scan the ship	Ancient shipwreck
inspect_pipeline_pool	Inspect the pipeline	Industrial pool with pipelines
inspect_pipeline_sea	Inspect the pipeline	Subsea pipeline

Tracking

Task Code	Instruction	Scenario
follow_boat	Follow the boat	Open sea

Transporting

Task Code	Instruction	Scenario
transfer_red_shallow	Pick up the red cylinder and transfer it to the box	Shallow ocean

Data Statistics

Overall

Metric	Train	Test	Total
Episodes	1,750	525	2,275
Frames	696,990	208,605	905,595
Videos	3,500	1,050	4,550

Per-Task Breakdown

Task	Train Episodes	Train Frames	Test Episodes	Test Frames
follow_boat	50	18,061	15	5,026
goto_charge_station	100	13,371	30	4,437
goto_water_tower	100	29,505	30	9,084
inspect_pipeline_pool	50	29,609	15	8,828
inspect_pipeline_sea	50	33,884	15	10,156
pick_blue_factory	100	38,038	30	11,857
pick_blue_shallow	100	35,953	30	11,371
pick_bluex_factory	100	38,461	30	11,505
pick_bluex_shallow	100	38,486	30	10,843
pick_pipe0_factory	100	38,683	30	10,942
pick_pipe0_shallow	100	37,205	30	11,411
pick_pipe1_factory	100	36,997	30	11,113
pick_pipe1_shallow	100	37,025	30	10,963
pick_red_factory	100	37,829	30	11,645
pick_red_shallow	100	36,914	30	10,990
pick_redx_factory	100	38,455	30	11,433
pick_redx_shallow	100	36,428	30	10,398
scan_ship_ancient	50	37,046	15	11,008
scan_ship_modern	50	33,868	15	10,285
transfer_red_shallow	100	51,172	30	15,310
Total	1,750	696,990	525	208,605

Data Schema

Both train/ and test/ follow the LeRobot v2.1 format. Each episode is stored as a Parquet file with the following features:

Observation

Feature	Dtype	Shape	Description
observation.images.ego	video	(240, 320, 3)	Front-facing ego camera RGB video
observation.images.wrist	video	(240, 320, 3)	Wrist-mounted end-effector camera RGB video
observation.state	float32	(29,)	Robot proprioceptive state vector

State Vector Breakdown (29-dim)

Component	Indices	Dim	Description
joint_pos	0–5	6	Arm joint positions
pwm	5–13	8	Thruster PWM values
joint_v	13–18	5	Arm joint velocities
dvl_v	18–21	3	Doppler Velocity Log velocity
imu_av	21–24	3	IMU angular velocity
imu_la	24–27	3	IMU linear acceleration
pressure	27–28	1	Pressure sensor reading
dvl_h	28–29	1	DVL altitude

Action

Feature	Dtype	Shape	Description
action	float32	(13,)	Robot action command

Action Breakdown (13-dim)

Component	Indices	Dim	Description
joint_pos	0–5	6	Arm target joint positions
pwm	5–13	8	Thruster PWM commands

Additional Features

Feature	Dtype	Shape	Description
target_pos	float32	(6,)	Target pose in robot local frame (x, y, z, roll, pitch, yaw)
timestamp	float32	(1,)	Frame timestamp in seconds
frame_index	int64	(1,)	Frame index within episode
episode_index	int64	(1,)	Episode index
index	int64	(1,)	Global frame index
task_index	int64	(1,)	Task index (maps to tasks.jsonl)

Video Metadata

Property	Value
Resolution	240 × 320
Codec	AV1
Pixel Format	YUV420P
FPS	10
Channels	3 (RGB)
Audio	No

Loading the Dataset

Using LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load the training split
train_dataset = LeRobotDataset("Vincent2025hello/usim", root="train")

# Load the test split
test_dataset = LeRobotDataset("Vincent2025hello/usim", root="test")

# Iterate through episodes
for episode in train_dataset:
    ego_image = episode["observation.images.ego"]       # (240, 320, 3) numpy array
    wrist_image = episode["observation.images.wrist"]   # (240, 320, 3) numpy array
    state = episode["observation.state"]                # (29,) numpy array
    action = episode["action"]                          # (13,) numpy array
    task_index = episode["task_index"]                  # scalar
    print(f"Task: {train_dataset.meta.tasks[task_index]}")

Using Hugging Face Datasets

from datasets import load_dataset

# Load from the repository
dataset = load_dataset("Vincent2025hello/usim")

Citation

If you use this dataset in your research, please cite:

@misc{gu2025usimu0visionlanguageactiondataset,
      title={USIM and U0: A Vision-Language-Action Dataset and Model for General Underwater Robots}, 
      author={Junwen Gu and Zhiheng Wu and Pengxuan Si and Shuang Qiu and Yukai Feng and Luoyang Sun and Laien Luo and Lianyi Yu and Jian Wang and Zhengxing Wu},
      year={2025},
      eprint={2510.07869},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.07869}, 
}

License

This dataset is released under the Apache 2.0 License.

Acknowledgements

• Stonefish — Physics-based underwater simulator
• stonefish_ros — ROS interface for Stonefish
• LeRobot — Dataset format and loading utilities