Abstract
We present StepNav, an efficient planning framework for visual navigation that generates reliable trajectories using structured trajectory priors. Unlike existing methods that rely on unstructured noise, StepNav leverages multi-modal trajectory initialization combined with conditional flow matching for efficient and safe path generation. Our approach incorporates a success probability field that predicts safe navigation regions and a dynamics-inspired feature projection module that enforces physical continuity in spatiotemporal features. Experiments demonstrate that StepNav achieves superior navigation performance while maintaining real-time efficiency suitable for autonomous systems.
Pipeline of StepNav: structured trajectory priors guide efficient trajectory generation.
Visualization of the success probability field predicting safe navigation regions.
Comparison of trajectory generation results with different methods.
Navigation Demonstrations
BibTeX
@misc{luo2026stepnavstructuredtrajectorypriors,
title={StepNav: Structured Trajectory Priors for Efficient and Multimodal Visual Navigation},
author={Xubo Luo and Aodi Wu and Haodong Han and Xue Wan and Wei Zhang and Leizheng Shu and Ruisuo Wang},
year={2026},
eprint={2602.02590},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2602.02590},
}