XR-Lab’s project was selected to join the Undergraduates Pursuing Research in Science and Engineering (UPRISE) Program
UNDERGRADUATE SUMMER RESEARCH PROGRAM IN SCIENCE AND ENGINEERING
May 6 – July 26, 2024
Project title: Reinforcement Learning (RL) system in Game Engine
Student: Mikhail Nikolaenko. UC. PI: Ming Tang.
This project proposes the development of a sophisticated reinforcement learning (RL) system utilizing the robust and versatile environment of Unreal Engine 5 (UE5). The primary objective is to create a flexible and highly realistic simulation platform that can model a multitude of real-life scenarios, ranging from object recognition, urban navigation to emergency response strategies. This platform aims to significantly advance the capabilities of RL algorithms by exposing them to complex, diverse, and dynamically changing environments. Leveraging the advanced graphical and physical simulation capabilities of UE5, the project will focus on creating detailed and varied scenarios in which RL algorithms can be trained and tested. These scenarios will include, but not be limited to, urban traffic systems, natural disaster simulations, and public safety response models. The realism and intricacy of UE5’s environment will provide a challenging and rich training ground for RL models, allowing them to learn and adapt to unpredictable variables akin to those in the real world.
Synthetic material for AI training
Using a game engine to generate synthetic training data offers significant advantages for AI training in image classification, object detection, and animation classification. Synthetic data, created in controlled virtual environments, allows for generating large, diverse, and perfectly labeled datasets. This contrasts with human-labeled material, which is often expensive, time-consuming, and prone to errors. Synthetic data can be tailored to specific needs, ensuring comprehensive coverage of various scenarios and edge cases that might be underrepresented in real-world data. Additionally, synthetic environments facilitate the manipulation of variables, such as lighting and angles, providing a robust training ground for AI models. Overall, the use of synthetic material enhances the efficiency and accuracy of AI training compared to traditional human-labeled datasets.
Example
The Roof AI Training project focuses on training an AI system using Unreal Engine 5 (UE5) to generate synthetic data for machine learning applications, specifically for detecting damaged asphalt shingles on roofs. It highlights the need for massive labeled datasets to train AI effectively and discusses the role of Convolutional Neural Networks (CNNs) in image recognition tasks. UE5 is utilized to create procedurally generated roofing textures and environments, providing a cost-effective and scalable solution for data generation. The research project involves generating high-quality synthetic images of roofs, including damaged areas, to train and evaluate a CNN model. The AI aspect includes using Simplex Noise and 3-tab pattern algorithms for realistic texture generation, followed by training a CNN using PyTorch on the generated data. Preliminary results indicate that the model accurately detects most missing shingles, though further improvements are needed in bounding box precision and detection accuracy. Future goals include enhancing the procedural generation algorithm, optimizing the CNN architecture for better performance, and iteratively refining the AI through continuous training and evaluation using both synthetic and real-world data.
Related projects: Building Safety Analysis with Machine Learning
