aiqbench analysis
StageVLN addresses a critical bottleneck in Vision-and-Language Navigation (VLN) by decoupling complex geometric reasoning from the deployment phase. By utilizing frozen geometry foundation models solely during training to guide hierarchical states, the framework achieves state-of-the-art results without incurring the computational overhead or architectural dependencies typically associated with explicit map encoders at inference time. This approach demonstrates how privileged spatial and trajectory supervision can enhance VLM-based agents while preserving efficiency for real-world robotics applications.
StageVLN: Spatial and Trajectory Auxiliary Guidance for Efficient Vision-Language Navigation
arXiv:2610.05664v1 Announce Type: cross Abstract: Vision-and-Language Navigation (VLN) policies increasingly benefit from strong semantic priors provided by large vision-language models (VLMs). However, standard action supervision does not explicitly encourage intermediate representations to preserve scene geometry, relative orientation, or global episode progress. Incorporating depth estimators, explicit maps, point clouds, or geometry foundation models at inference can provide such structure but introduces additional computation, memory overhead, and architectural dependence during deployment. We introduce StageVLN, a training framework that shapes navigation representations through privileged spatial and trajectory guidance while preserving the original inference pathway. A frozen geometry foundation model provides multi-level spatial guidance to hierarchical navigator states, while relative-heading and expert-route progress objectives provide complementary trajectory-state supervision. All auxiliary components are used only during training and removed at deployment. On R2R-CE validation-unseen, StageVLN achieves 56.3\% SR and 51.4\% SPL with a 4B-parameter backbone, without an additional geometry encoder at inference. On RxR-CE, it achieves 54.3\% SR without additional navigation training data or a geometry encoder at inference.
Abstract:Vision-and-Language Navigation (VLN) policies increasingly benefit from strong semantic priors provided by large vision-language models (VLMs). However, standard action supervision does not explicitly encourage intermediate representations to preserve scene geometry, relative orientation, or global episode progress. Incorporating depth estimators, explicit maps, point clouds, or geometry foundation models at inference can provide such structure but introduces additional computation, memory overhead, and architectural dependence during deployment. We introduce StageVLN, a training framework that shapes navigation representations through privileged spatial and trajectory guidance while preserving the original inference pathway. A frozen geometry foundation model provides multi-level spatial guidance to hierarchical navigator states, while relative-heading and expert-route progress objectives provide complementary trajectory-state supervision. All auxiliary components are used only during training and removed at deployment. On R2R-CE validation-unseen, StageVLN achieves 56.3\% SR and 51.4\% SPL with a 4B-parameter backbone, without an additional geometry encoder at inference. On RxR-CE, it achieves 54.3\% SR without additional navigation training data or a geometry encoder at inference.
Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
arXiv:2610.05664 [cs.CV]
(or arXiv:2610.05664v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2610.05664 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 258 words
Related coverage
Pointing-VLA: Typed Spatial Grounding Interfaces for Vision-Language-Action Manipulation
arXiv · Aug 25, 2026
US-VLA: An Ultrasound Vision-Language-Action Model for Embodied Abdomina
arXiv · Aug 18, 2026
Latent-Centroid Steering: Single-Pass Classifier-Free Guidance for Command-Aligned Autonomous Driving
arXiv · Aug 4, 2026
Universal Pose Pretraining for Generalizable Vision-Language-Action Policies
arXiv · Sep 30, 2026