AquaWorld: Structure-Consistent Underwater World Generation for Robot Simulation

Bin Peng; Tiandong Zhang; Ruidong Wang; Min Luo; Shuo Wang

cs.ROSep 19, 2026

AquaWorld: Structure-Consistent Underwater World Generation for Robot Simulation

Authors: Bin Peng, Tiandong Zhang, Ruidong Wang, Min Luo, Shuo Wang

Organizations: State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · Ocean College, Zhejiang University, Zhejiang, China

Abstract

Underwater robot simulation requires diverse environments in which terrain, scene composition, tasks, and currents remain mutually consistent. We present AquaWorld, a world-generation framework that preserves these relationships through shared terrain structure. A language-conditioned plan generates 3D terrain and shared structural references that guide asset placement, task definition, and inflow specification under stochastic variation. The framework incorporates over 10,000 underwater-compatible assets, predicts reusable terrain-conditioned mean-flow fields through a CFD-supervised residual model, and supports conventional underwater vehicles and bio-inspired robotic fish. On 24 paired terrains, structure-consistent randomization produces substantially better cross-factor consistency than independent randomization. In a matched-budget policy-training comparison, structurally coherent randomization achieves a validation success rate 21% higher than independent randomization. In separate physical experiments, a simulation-trained visual navigation policy succeeds in 95% of physical tank trials without updating its perception or control modules. Overall, AquaWorld provides a practical way to generate varied underwater environments while retaining the structural relationships needed for flow simulation and robot learning.

Figures & tables

Framework	World construction	Language control	Terrain-linked outputs	Flow modeling	Robot models
Infinigen [ 7 ]	Procedural	–	Assets	–	–
DreamSea [ 15 ]	Learned	–	–	–	–
OysterSim [ 11 ]	Asset synthesis	–	–	–	–
ChatSim [ 8 ]	Scene editing	$\checkmark$	–	–	–
OceanSim [ 6 ]	Imported	–	–	–	Single-body
HoloOcean [ 3 , 9 ]	Image-conditioned	–	Assets	Prescribed	Single-body

TABLE I: Comparison of representative underwater world-generation and simulation frameworks.

Fig. 2: Language-conditioned terrain generation and structure-aware scene instantiation. A validated ScenePlan selects terrain operators to generate 3D geometry and structural references, which are then used for scene and task instantiation.

Fig. 3: Terrain-conditioned mean-flow generation. Terrain features and resolved inflow conditions drive a residual encoder–decoder. Regional CFD supervision distinguishes the full field, near-wall and task regions, wake and reverse-flow regions, and upper and far downstream regions. The illustrated slices compare reference and predicted mean velocities.

Fig. 4: Representative robot morphologies supported in AquaWorld. The simulation includes conventional underwater vehicles and articulated bio-inspired robots with multi-link bodies or discretized undulating fins.

Method	Reg.	Align.	Full (%)	Near (%)	Task (%)	Wake RMSE/ $U$
Global supervision	–	$\checkmark$	4.48	13.08	8.07	0.238
Regional supervision	$\checkmark$	–	4.52	12.53	7.33	0.225
Full model	$\checkmark$	$\checkmark$	4.42	12.02	6.93	0.223

TABLE II: Ablation of regional supervision and inflow alignment on the development set.

Fig. 5: Visual navigation and sim-to-real transfer. (a) Target- and obstacle-perception CNNs process onboard RGB observations, followed by an MLP PPO policy for robot control. (b) Representative runs in the simulated tank and the corresponding physical setup, showing overhead trajectories and onboard camera views during target approach and obstacle avoidance.

Fig. 6: Matched world-generation comparison. Independent and structurally coherent randomization use the same terrain and matched randomized resources. AquaWorld retains terrain-relative scene, flow, and task associations through shared structural references. Numerical annotations in the image describe the displayed example.

Metric	Independent	AquaWorld
Flow-direction error $\downarrow$	$31.41^{\circ}$	$\mathbf{6.68^{\circ}}$
Asset–support agreement $\uparrow$	$50.65\%$	$\mathbf{98.00\%}$
Task–channel coverage $\uparrow$	$73.25\%$	$\mathbf{100.00\%}$

TABLE III: Cross-factor consistency over 24 matched terrain pairs.

Robotics Simulation

Robotics simulation builds virtual robots and environments to model, train, and evaluate autonomous systems before real-world deployment.

See all 320 papers

Coastal Environment Generation with HoloOcean
Sep 9, 2026Abigail Austin, Brady Moon, Joshua G. MangelsonRobotics Simulation 3D Scene Generation
Brigham Young University
D3D-GEN: Robot-Aware Domain-Grounded Interactive 3D World Generation for Social Robotics
Aug 12, 2026Anh Duc Do, Volodymyr Shcherbyna, Tai Duc Nguyen +5Robotics Simulation Functional 3D Scene Generation
Singapore Management University · Max Planck Institute · Technical University Braunschweig
EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI
Jul 8, 2026Xinjie Wang, Liu Liu, Taojun Ding +93D Asset Generation Robotics Simulation
Horizon Robotics · WuwenAI

Sep 9, 2026J/K move · Enter open · S save

Coastal Environment Generation with HoloOcean

Abigail Austin, Brady Moon, Joshua G. Mangelson

Brigham Young University

Robotics Simulation 3D Scene Generation Procedural Content Generation

Open paper AskPDF arXiv

Marine robotic simulation provides a safe and inexpensive method of developing and testing algorithms for unmanned underwater vehicle (UUV) and unmanned surface vessel (USV) autonomy and perception before full field deployment. However, these simulations are often limited by the availability of simulated environments. Current marine robotics simulation suites offer manual ways to edit or create environments, but they require existing data or specialized knowledge of the environment system. To address these issues, we introduce a novel Unreal Engine 5 level generation pipeline that enables automatic creation of coastal environments for HoloOcean. Our pipeline relies on a user-provided overhead image of a coastal scene. The pipeline then uses the image to generate height map data, as well as automatically select assets and place them in the environment.

Abstract

Robotics Simulation

Explore similar work

Coastal Environment Generation with HoloOcean

D3D-GEN: Robot-Aware Domain-Grounded Interactive 3D World Generation for Social Robotics

EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI