Learning Semantic Navigation Primitives for Human Habitats Using a Recurrent State-Space World Model
Published in WRC 2026 SARA, 2026
Complex semantic navigation tasks can be decomposed into reusable behaviors. In this work, we learn two Semantic Navigation Primitives—moving to the middle of a room and passing through a door—with the DreamerV3 recurrent state-space world model, then compose them to navigate between rooms in human habitats.
The models are trained with modest commercial compute and deployed entirely on an NVIDIA Jetson AGX Orin without cloud access. Individual primitives achieve a peak success rate of 94%, demonstrating a path toward resource-efficient, privacy-preserving semantic navigation on embedded robot hardware.
Conference program · Presentation videos
The IEEE Xplore record is not yet available and will be added when published.
Recommended citation: A. Elksnis, Z. Luo, N. Wang, and M. Pearson (2026). "Learning Semantic Navigation Primitives for Human Habitats Using a Recurrent State-Space World Model." World Robot Conference 2026, Symposium on Advanced Robotics and Automation.
Download Paper