Author ORCID Identifier

https://orcid.org/0009-0009-6967-0364

Date of Award

Spring 5-15-2025

Document Type

Thesis (Master's)

Department or Program

Computer Science

First Advisor

Alberto Quattrini Li

Second Advisor

Soroush Vosoughi

Third Advisor

James Mahoney

Abstract

From self-driving cars navigating city streets to all-terrain vehicles tackling rugged landscapes, recent leaps in robotic autonomy due to fast pace development in deep learning are reshaping how machines interact with the real world. However, autonomy in the aquatic environment is still limited, due to difficulty in testing and unavailability of realistic simulation environments.

In this project, we aim to create an automated system that simplifies the processes of creating synthetic datasets for marine robots navigation training tasks. We achieved this through a land cover map controlled terrain generation. Our goal is to provide an automatic terrain generation system that includes multiple terrain types, give users a certain level of customizability and reach a realistic level for simulation.

Currently, obtaining real-world datasets for marine robots is high-cost and difficult. Datasets that are currently in use are either hand captured by humans or specific types of robots, which will restrict the usability to limited sensor types. Additionally, the equipment used in the capturing process is expensive and subject to risk of damage. Finally, the time consuming nature of such tasks also limited the creation of large datasets.

To generate synthetic, but realistic datasets, we propose an approach that follows two steps: 1) generate the topology of the environment and 2) create a real 3D world asset that can be simulated with a realistic physics game engine. For generating the topology of the environment, we provide the environment topology generation in two modes: a) given an input map, we trained U-net model to output a classified land-cover map; b) without any map, for generating random terrains that are realistic, we propose a diffusion-model. The Diffusion-model will output a randomly generated map which is similar to the land-cover map. The user can use either map as input to generate a terrain.

For generating realistic 3D worlds, we propose a PCG system that takes as input the land-cover at the previous step and generates terrains in a game engine. We chose UE5 as our main tool given its realism, available in literature, and the availability of routines that can be used for PCG.

Download

Share

COinS