Model-Based Inference and Deep Representations in Concert for Robot Autonomy

This talk explores how model-based inference and deep representations can be composed to enable robust robot autonomy, with a focus on perception and decision making in complex environments. Building on some of the recent research activities and achievements of the Laboratory for Autonomous Systems and Mobile Robotics (LAMOR) in visual odometry, monocular depth estimation, and heterogeneous multi-sensor calibration, we will show how a modular pipeline combining geometry, probabilistic estimation, and transformer-based architecture improves accuracy, robustness, and generalization of visual ego-motion estimation methods. On the decision-making side, we will discuss an approach to high-dimensional motion planning for efficient navigation of robotic manipulators in cluttered scenes, finishing with a take on recent results in active visual attention in autonomous systems, thus closing the loop between perception, estimation, and planning.

Ivan Marković is a Full Professor at the University of Zagreb Faculty of Electrical Engineering and Computing, Croatia (UNIZG-FER) where he leads the Robotic perception and sensor fusion group within the Laboratory for Autonomous Systems and Mobile Robotics – LAMOR (http://lamor.fer.hr).