Abstract: We are seeing embedded-scale, compact radio frequency devices emerge for a variety of portable sensing applications. For example, a millimeter wave radar on a phone can recognize gestures in no/low light conditions. However, due to fundamental limitations, their use has been limited to coarse-grained sensing tasks - as opposed to full fledged imaging. In this talk, I will present techniques that achieve high quality sensing through occlusions paving the way for new applications in cyber-physical systems, robotics and other autonomous systems.

 

First, I will show an imaging technique for imaging automotive tires (even in dusty, debris-prone conditions) at high quality. This has created valuable automotive health sensing infrastructure for use in agricultural, mining and passenger vehicles. Second, I will talk about machine learning based super-resolution, learning from expensive sensors to train compact, cheaper sensors to perceive robustly in harsh conditions. I show how this has enabled a firefighting robot to perform search and rescue operations through camera-denied, smoky environments. Finally, I will introduce neural radiance field techniques to generate realistic, synthetic radio frequency data to further bootstrap large data needed to train these large models.

 

I will conclude the talk with my future vision on how to push towards camera-quality perception from these devices – from both a computational and hardware perspective. Higher fidelity radio frequency perception and communication will unlock creative exploration of applications, seeing through new scattering and occluded media and solving pressing issues.

 

Short Bio: Akarsh Prabhakara is a Ph.D. candidate in Electrical and Computer Engineering at Carnegie Mellon University. His research interests lie at the intersection of wireless systems and cyber-physical systems. He has won Best Paper Honorable Mention at MobiSys, GetMobile Research Highlight twice, and loves to demonstrate his systems end-to-end at conferences winning Best Demo winner and runner-up at MobiSys, IPSN and MobiCom. His work on imaging automotive tires has been adopted by Bridgestone, a global tire manufacturer.