Bonjour à tous,

Vous êtes tous et toutes cordialement invité.es à assister à la présentation de projet du prédoc III de Miguel Saavedra-Ruiz, le 14 août à 13h (mode hybride).

Titre : Static... But Not Quite: Towards Spatio-Temporal World Modeling of Semi-Static Environments in Robotics

Date: jeudi 14 juin à 13h.

Location: Auditorium 1, Mila, 6650 (2e étage)

Jury

Résumé

World representations are a cornerstone of robotics research as they allow the state of the world to be encoded in a computer program and used for downstream tasks. However, in the past two decades, most research efforts have overlooked a crucial aspect of the real world: its dynamic nature. As a result, state-of-the-art world representations are brittle in the presence of dynamics and struggle to model the rich, multimodal dynamics of real-world changes.

This report explores two approaches to modeling changes in dynamic environments, with a focus on semi-static settings, whereelements appear static but change over time. We begin with the harmonicexponential filter (HEF), a Bayesian filter based on noncommutative harmonic analysis that can model arbitrary multimodal distributions. Then we present Perpetua, a method that estimates when features may appear or disappear in semi-static environments while maintaining multiple hypotheses over their state dynamics. Lastly, we outline three research directions for dynamic world representations in robotics. First, to bridge the gap between learning-based and robotics approaches, we propose collecting a months-long dataset capturing semi-static dynamics in a large indoor environment. Second, we discuss an extension of Perpetua that can model spatio-temporal dynamics using deep generative methods. Third, we explore the integration of persistence estimators into state-of-the-art world representations, aiming to extend their applicability to more realistic and challenging scenarios.