Autonomous systems, including self-driving cars and unmanned aircraft, rely on modeling and simulation for their development. However, the training process for these systems can take extensive periods and often fails to account for real-world uncertainties, a challenge known as the simulation-to-real gap.
To address this issue, the Defense Advanced Research Projects Agency (DARPA) has initiated the Transfer Learning from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program. As part of this initiative, DARPA recently awarded a $1.2 million grant to University of Central Florida (UCF) researchers George Atia and Yue Wang for their project, “Distributionally Robust Approaches to Transfer Learning.”
Atia, an associate professor in the Department of Electrical and Computer Engineering, expressed his appreciation for the recognition and opportunity presented by the DARPA award, highlighting the competitive nature of the funding environment. Over the next 18 months, Atia and Wang will focus on developing artificial intelligence technologies to help autonomous systems better adapt to unknown variables. Despite the complexity and realism of current simulation environments, they often do not prepare systems for unexpected changes, such as variations in flight dynamics or lighting conditions for drones transitioning from urban to coastal areas.
Additionally, the training speed is a significant challenge, as it can take millions of simulated episodes over several years for an autonomous system to be adequately prepared for real-world applications. The UCF team aims to address this by designing technology that can expedite the training process.
Atia elaborated on their approach, comparing it to teaching a robot to navigate a busy city after mastering a simple maze. Their method seeks to equip robots with the ability to handle surprises, thus enabling faster learning and better performance. This approach bridges theoretical knowledge with practical application, improving knowledge transfer effectiveness, especially in scenarios with limited real-world data.
The potential applications of their research extend beyond defense. In healthcare, their robust knowledge transfer methods could enhance personalized care by facilitating the transfer of treatment plans between patients. Similarly, decision-making policies tailored for specific road conditions could be adapted to other environments, improving safety and efficiency in autonomous driving. Atia emphasized that by addressing the limitations of traditional machine learning methods, their research could revolutionize various industries and enable transformative solutions to complex problems.
Photo credit Antoine Hart, University of Central Florida.
