Year: 2023
Type: article
Abstract: An autonomous system that cannot represent how confident it is in its own perception, localization, and predictions cannot be trusted to act safely in a partially observed world, and probabilistic graphical models remain the principal mathematical language for making that confidence explicit. This article synthesizes the literature on uncertainty quantification for autonomous systems across two historically separate traditions: classical Bayesian filtering and factor-graph SLAM, built on Kalman's original recursive estimator and matured through particle filtering and modern sparse smoothing, and the newer body of work retrofitting principled probabilistic reasoning onto deep neural networks, from Bayesian neural networks and Monte Carlo dropout through deep ensembles and single-pass evidential methods. We trace how both traditions converge on the same underlying representational problem, distinguishing aleatoric from epistemic uncertainty and propagating a calibrated belief rather than a point estimate, and examine how that belief is used for planning under uncertainty via partially observable Markov decision processes, and applied concretely in object detection and multimodal trajectory prediction for autonomous driving. Read together, the evidence supports probabilistic graphical models as a coherent, empirically validated foundation for uncertainty-aware autonomy, while leaving calibration under distribution shift, uncertainty propagation across module boundaries, and the computational cost of principled Bayesian methods on embedded hardware as an open research agenda.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7213623229