Building trust in autonomous systems with an AI framework for privacy, safety, and reliability in data, software, and robotics
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Subject
Abstract
In this research, we propose an innovative framework designed to enhance privacy, safety, and reliability in data science, software development, and AI-empowered robotics, with a primary focus on building trust in autonomous systems. Our approach initially emphasizes the application of deep learning techniques to recover potential data values while bolstering data protection through privacy-preserving methods such as Differential Privacy. This is applied to heterogeneous data sources, including Census Demographics, OpenData 311 Calls, Neighborhood Crime, and LandBank data. As a result, we have achieved high accuracy in recovering missing and invalid data, while successfully preserving overall data distribution and ensuring privacy. For the second objective, our focus shifts to improving software system security through anonymous data collection tools. These tools are further enhanced with explainable and predictable measures of trust in human-machine collaboration, ensuring that these advanced processes remain transparent, accessible, and interpretable for all users, thereby increasing their overall trustworthiness. Within autonomous systems, robotics has been considered the most complex due to its continuous integration of data, software, and hardware. Because of this, privacy, safety, and reliability are the ultimate factors in deciding its trustworthiness as it involves not only a real-time decision-making process but also quite often heavy human-robot collaboration. As an integral concept of AI-empowered robotics, digital twins revolutionize the collaboration between virtual and real robotic systems, alongside reducing risks in reality and fostering human-robot interactions. Besides achieving state-of-the-art results in Object Goal Navigation, our approach emphasizes the use of digital twins, reinforcement learning, and imitation learning in other domains such as healthcare, surgery, and emergency handling, highlighting the importance of safety and reliability in the field of AI-empowered robotics. To gain trust between humans and autonomous systems, we subsequently deploy the diversely trained models in mixed-reality environments (AR/VR/XR), allowing experts to engage in an immersive and interactive experience where they can closely evaluate the robots' behaviors and decision-making processes in realistic scenarios.
Table of Contents
Introduction -- Data collection and integration with boundary assignment -- Big data analytics framework for predictive analytics using public data with privacy preserving -- AI-powered data collection of unfiltered opinions -- Dynamic form generation and chart visualization -- Optimizing robotic systems quality through virtual-real integration with advanced reinforcement learning -- Enhancing reliability in critical task-handling with digital twins -- Evaluating safety and reliability in object goal navigation tasks -- Conclusion and future work
DOI
PubMed ID
Degree
Ph.D. (Doctor of Philosophy)