Uncertainty-Aware Stochastic Hybrid World Models with Neural Map Memory for Autonomous Navigation in Partially Observable Grid Environments
Computational Discovery and Intelligent Systems (CDIS) · 25.04.2026 · 3 (1) · 38-56
DOI: 10.66279/t95nh857 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computational Discovery and Intelligent Systems (CDIS), online erschienen am 25.04.2026. Verfasst von Arwa Saad und 2 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Autonomous navigation in partially observable environments remains a significant challenge for reinforcement learning agents due to incomplete observations, stochastic dynamics, and uncertainty in spatial perception. World models are robust at learning how the environment changes over time, and neural map architectures are competitive at representing spatial memory. However, most current methods treat these parts separately and don't often include explicit uncertainty estimation, thereby reducing navigation reliability and exploration efficiency. This paper introduces SHWM-NM (Stochastic Hybrid World Model with Uncertainty-Aware Neural Map Memory), a unified framework that integrates stochastic latent dynamics modeling, structured neural map memory, and multi-level uncertainty estimation to enhance autonomous navigation capabilities. The proposed architecture combines a stochastic hybrid world model with an uncertainty-aware neural map that explicitly represents spatial information and associated uncertainty. A policy learning module then employs these estimates to help with exploration and decision-making when not all information is available. When tested on MiniGrid-based navigation tasks, SHWM-NM significantly outperforms deterministic world model baselines. The proposed framework increases the average reward from 1.55 to 4.61, raises the success rate from 15% to 46%, and reduces the average trajectory length from 42.5 to 30.7 steps. Also, epistemic uncertainty decreased from 0.0073 to 0.0017 during training, reflecting a modest but consistent improvement in model confidence. During training, which means indicating improved modeling of environment dynamics. These results show that modeling stochastic dynamics, spatial memory, and uncertainty together in a single architecture demonstrates strong performance. This is a promising approach to making promising decisions and getting around in environments that aren't fully observable.
Zitieren
APA
Saad, A., Ibrahim, A., & Gupta, S. (2026). Uncertainty-Aware Stochastic Hybrid World Models with Neural Map Memory for Autonomous Navigation in Partially Observable Grid Environments. Computational Discovery and Intelligent Systems (CDIS), 3(1), 38-56. https://doi.org/10.66279/t95nh857
Vancouver
Saad A, Ibrahim A, Gupta S. Uncertainty-Aware Stochastic Hybrid World Models with Neural Map Memory for Autonomous Navigation in Partially Observable Grid Environments. Computational Discovery and Intelligent Systems (CDIS). 2026;3(1):38-56. doi:10.66279/t95nh857
BibTeX
@article{Saad2026,
author = {Saad, Arwa and Ibrahim, Amer and Gupta, Shashi},
title = {Uncertainty-Aware Stochastic Hybrid World Models with Neural Map Memory for Autonomous Navigation in Partially Observable Grid Environments},
journal = {Computational Discovery and Intelligent Systems (CDIS)},
year = {2026},
volume = {3},
number = {1},
pages = {38--56},
doi = {10.66279/t95nh857}
}
EUVIMED