Causal Representation of the Simultaneous Localization And Mapping Problem - LAAS-Robotique
Pré-Publication, Document De Travail Année : 2024

Causal Representation of the Simultaneous Localization And Mapping Problem

Résumé

We address the problem of spatial inference in robotics, commonly known as “Simultaneous Localization And Mapping” (SLAM), through the lens of causal inference. State-of-the-art approaches generally proceed by feeding the SLAM system with noisy sensor measurements and statistical assumptions, so as to get the joint posterior probability distribution of the decision variables conditioned on the measurements. Probabilistic inference techniques are used to produce the best posterior estimate. In contrast, in the proposed approach, the estimation of the decision variables is conducted on a scheme which makes use of causal assumptions. Since understanding the distinction between the probabilistic and the causal paradigms is paramount in this work, a doctrinal synthesis is provided, supported by an historical analysis. In particular, it is reminded that probability theory pioneers like Laplace handled uncertainty within a quasi-deterministic framework. Such a conception found an apropos application in the geodetic adjustment problem (triangulations for cartography), which shares many structural commonalities with SLAM. Using modern causal inference tools, in particular causal graphs and the structural causal model (SCM) championed by Pearl and co-authors, we develop a causal view of the SLAM problem in order to address some of the longstanding issues. We show how causal hypotheses can be articulated, before appropriately exploiting the conditional independences exhibited by the model. As a result, the adjustment problem boils down to a constrained least squares optimization problem. The overall framework is referred as CARLIT (Causal Approach to Represent Locomotion and Internalize Topography). The conceptual ramifications are analyzed, e.g., concerning the processing of loop-closure events or the incorporation of landmarks. While toy SLAM problems are used to showcase the approach, various evaluations are also provided on synthetic datasets as well as a real-scale experimentation.
Fichier principal
Vignette du fichier
monograph-tari-danes-v0.pdf (28.66 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
licence
Copyright (Tous droits réservés)

Dates et versions

hal-04860487 , version 1 (31-12-2024)

Licence

Copyright (Tous droits réservés)

Identifiants

  • HAL Id : hal-04860487 , version 1

Citer

Joël Tari, Patrick Danès. Causal Representation of the Simultaneous Localization And Mapping Problem. 2024. ⟨hal-04860487⟩
0 Consultations
0 Téléchargements

Partager

More