Cornelia Fermüller (University of Maryland), “Computational Principles of Embodied Intelligence for Robust Motion Perception and Action”

Abstract
Understanding the computational principles of embodied intelligence is central to advancing robotic systems that perceive and act in complex environments. This talk explores key principles—low power consumption, robustness, and generalizability—as they emerge in the context of motion perception and action. For visual navigation, evidence is presented that challenges the conventional SLAM paradigm, which relies on correspondence estimation and 3D scene reconstruction. Instead, 3D motion estimation and scene segmentation can be achieved using 1D normal flow measurements derived from image gradients, offering a simpler and more robust alternative. The effectiveness of this approach is demonstrated through implementations on drones equipped with both standard and neuromorphic dynamic vision sensors. Further, it is shown that physical interaction tasks do not necessarily require explicit depth estimation; rather, distance can be inferred in action-dependent units grounded in control dynamics. Finally, the role of visual motion in action understanding is examined, focusing on how motion-derived primitives support robust and generalizable representations of action, opening new avenues for embodied intelligence in robotic systems.

This talk is part of Aravind Battaje‘s course “Mind, Body, Environment: An Interactive Seminar on Embodied Intelligence,” a seminar introducing to key theories and research highlighting this shift in perspective through invited lectures from experts in the field and interactive sessions.

Bio

Cornelia Fermüller is a Research Scientist at the University of Maryland’s Institute for Advanced Computer Studies (UMIACS), where she co-founded the Autonomy Cognition and Robotics (ARC) Lab and co-leads the Perception and Robotics Group. Her research lies at the intersection of computer vision, robotics, and human vision, with a focus on biologically inspired solutions for active vision systems. She has made significant contributions to the understanding of visual perception by developing computational models for visual motion analysis, 3D motion and shape estimation, texture analysis, and action recognition, as well as integrating perception, action, and reasoning to enable cognitive robots to learn and interpret human manipulation actions.

Dr. Fermüller holds an M.S. from the University of Technology, Graz, and a Ph.D. in Applied Mathematics from the Technical University of Vienna. Her recent work emphasizes the use of event-based, bio-inspired sensors for robust motion perception in challenging environments, with applications ranging from fast motion perception for drones to autonomous driving in diverse lighting conditions. She is the principal investigator of an NSF-sponsored Science of Learning Center Network for Neuromorphic Engineering, co-organizes the Neuromorphic Engineering and Cognition Workshop, and has been recognized for her leadership in interdisciplinary research bridging computational modeling and psychophysical studies of human vision.

For those who are not in Berlin but would like to join virtually:
https://tu-berlin.zoom-x.de/j/69207754612?pwd=IKxoTdY3dQWccHpce2nA0IsNkNxPHu.1

Photo created with DALL-E by Maria Ott.