Skip to content
Book Open access

Long-Horizon Vision-Language Navigation by Memory Editing

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 65 references

Abstract

Vision-Language Navigation (VLN) requires an embodied agent to follow natural-language instructions and navigate in previously unseen environments. Despite recent progress in continuous multimodal navigation, long-horizon VLN remains challenging since it requires preserving useful history and maintaining decision consistency under partial observability. In this paper, we address long-horizon VLN by memory editing and propose a dual-memory framework, MemEdit-VLN. Specifically, MemEdit-VLN improves long-horizon VLN through Semantic Memory Consolidation, which compacts redundant semantic observations, and training-free Recallable Spatial Memory Selection, which preserves geometry-grounded spatial history for later navigation decisions. To better evaluate memory-intensive navigation, we construct the R2R Long-Horizon (R2R-LH) dataset from standard VLN-CE episodes, with longer horizons and stronger cross-stage dependencies. Experiments on R2R-LH and standard short-horizon VLN-CE benchmarks demonstrate that MemEdit-VLN outperforms state-of-the-art methods, achieving a remarkable 71.7% relative improvement in success rate for long-horizon VLN.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.