AI Paper Atlas

Research map / Vision-language and multimodal / Robot action and navigation

One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation

Zhuoyuan Yu, Yuxing Long, Du Wang, Junyang Wang, Hongwei Fan, Hao Dong

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Embodied multimodal LLM agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Scan2Instruction, which turns a single brief iPhone scan of a building into a training environment and automatically synthesizes instructions to adapt a language-guided navigation policy without human demonstrations.

Problem

Language-guided navigation policies (VLAs) degrade in new buildings as layout, appearance, and language shift, and the standard fix of teleoperating new demonstrations on site does not scale.

Approach

Replaces human demonstrations with a single brief iPhone scan that is compiled into a complete training environment, from which navigation instructions and trajectories are automatically synthesized for on-site adaptation.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
post-train RealNav with PRPT-KS on Scan2Instruction episodes instead of generic pretraining onlyRealNav · five scanned scenes, real robotsuccess rate: 0.25 -> 0.85Improved
use PRPT-KS on a single scan instead of fine-tuning on 20 human teleoperated trajectories per sceneRealNav · five scanned scenes, real robotsuccess rate: 0.49 -> 0.85Improved
use PRPT-KS adaptation versus human-demo SFT in never-scanned regionsRealNav · unscanned regions, real robotsuccess rate: 0.41 -> 0.82Improved
fine-tune on 20 human demonstrations per sceneRealNav and CorrectNav · five scanned scenes, real robotsuccess rate: 0.25 -> 0.49 (RealNav); 0.22 -> 0.48 (CorrectNav)Improved
add generic navigation pretraining only (no scene adaptation) for RealNavRealNav · Gaussian simulation, five scenessuccess rate: 0.23 vs external baselines 0.20-0.23No clear effect
add the behavior-cloning anchor (planner-reference replay loss) (ablation)RealNav + PRPT-KS · Gaussian simulation, five scenessuccess rate: 0.30 -> 0.85Improved
add the Pareto-utility advantage policy-gradient term (ablation)RealNav + PRPT-KS · Gaussian simulation, five scenessuccess rate: 0.81 -> 0.85Improved
add the Pareto-weighted preference loss (ablation)RealNav + PRPT-KS · Gaussian simulation, five scenessuccess rate: 0.80 -> 0.85Improved
add the policy-gradient term or the preference loss (ablation)RealNav + PRPT-KS · Gaussian simulation, five scenesnavigation error (m): 1.31 -> 1.42 (policy gradient); 1.32 -> 1.42 (preference)Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy