Distribution-Based Representations for Efficient Human-Robot Communication

Ahalya Prabhakar, Postdoctoral Fellow, Ecole Polytechnique Federale de Lausanne (EPFL)

WHERE: Virtual

WHEN: March 14, 2022 12:30 pm-1:30 pmADD TO CALENDAR

WEB: 
Distribution-Based  Representations for Efficient Human-Robot Communication: Ahalya Prabhakar, Postdoctoral Fellow, Ecole Polytechnique Federale de Lausanne (EPFL)

ABSTRACT:
Efficient communication is necessary for real-world human-robot interaction. However, directly communicating potentially high-dimensional data about complex tasks can be intractable. In this talk, I show that robotic systems are capable of learning and acting on low-dimensional task representations in real-time. I introduce the use of spatial distributions as a concise representation of temporal trajectories to better describe tasks in human-robot collaboration and learning. In combination with information-theoretic controllers, I show that these spatial representations can be used in real-time control paradigms and result in successful task performance. In addition, I discuss how these motion representations can be used to design intuitive interfaces for real-time, dynamic human-swarm collaboration. Furthermore, I explore how these representations can be learned from data to facilitate robot learning, e.g., learning task representations directly from human demonstrations. I illustrate the efficacy of distribution-based representations for real-time task control using results from simulation examples and experiments, including field tests with a swarm of rovers.

BIO:
Ahalya is currently a postdoctoral fellow at EPFL with Professor Aude Billard. She received her PhD and Masters in Mechanical Engineering at Northwestern University with Professor Todd Murphey. She also received her B.Sc in Mechanical Engineering at Caltech. Her research explores methods for enabling efficient human-robot communication. Her current work focuses on combining learning and multimodal sensory fusion to enable safe, adaptive control in human-robot interactive settings.