NaviScale: Generating Large-Scale Semantic Map Datasets for Object Navigation

Chuanlin Lan; Yanwei Zheng; Yuxi Jing; Weijian Liu; Zhitong Zhou; Jiarui Fan; Fuzhen Zhuang; Xiao Zhang

cs.ROSep 23, 2026

NaviScale: Generating Large-Scale Semantic Map Datasets for Object Navigation

Authors: Chuanlin Lan, Yanwei Zheng, Yuxi Jing, Weijian Liu, Zhitong Zhou, Jiarui Fan, Fuzhen Zhuang, Xiao Zhang, +1 more

Organizations: School of Computer Science and Technology, Shandong University, Qingdao, China

Abstract

Embodied navigation requires spatial representations that generalize across unseen environments, yet collecting large amounts of annotated data from real 3D environments is difficult. We propose NaviScale for semantic-map-based object navigation (ObjectNav), whose predictor can be trained on pairs of partial and complete semantic maps without reconstructing a complete 3D environment for every training sample. The framework generates large-scale semantic map training data by composing floorplans of real homes with room-level semantic and obstacle maps extracted from MP3D and HM3DSem. NaviScale increases data diversity in two ways: inter-room scaling increases floorplan-level structural diversity, while intra-room scaling fills each fixed floorplan with different combinations of room maps matched by room category. Visibility through Ray Casting (VisRC) converts the composed maps into partial observations that account for field of view, sensing range, and occlusion. The resulting dataset contains 192,000 semantic maps generated from 24,000 floorplans associated with 12,794 properties. With 300k training iterations and the training and inference settings described in this paper, the system reaches 64.3% SR and 34.8% SPL on HM3D, together with 43.1% SR and 16.8% SPL on MP3D, without changing the prediction architecture. Additional experiments evaluate the quality of the composed maps, the effects of semantic-segmentation errors, and deployment on a physical robot.

Figures & tables

Fig. 1: Motivation and overview of NaviScale. (A) Real 3D scans capture real environments but are costly to scale, whereas procedural 3D environments are scalable but depend on predefined generation rules. (B) For semantic-map-based ObjectNav, the spatial predictor can be trained directly from partial and full semantic maps without constructing a complete 3D environment for every training sample. (C) NaviScale combines floorplans of real homes with room-level semantic and obstacle maps with matching room categories to generate large-scale training maps. (D) The resulting dataset contains 24,000 floorplans, 192,000 semantic maps, and 2.41 million room placements. Under the same prediction architecture and the same training and inference settings within each benchmark, replacing the original benchmark training data with NaviScale improves five-run mean SR from 59.6% to 64.3% on HM3D and from 40.3% to 43.1% on MP3D.

Fig. 2: Overview of NaviScale. Inter-room scaling collects distinct floorplans of real homes to vary global structure. Intra-room scaling assigns different semantic and obstacle room maps with matching room categories to each fixed layout, generating multiple full maps ( $S^{\text{full}},O^{\text{full}}$ ). VisRC then applies field-of-view, sensing-range, and occlusion constraints to produce partial observations ( $M_{t}$ ) for training.

Fig. 3: Ablation study results for the scaling dimensions under the 300k training schedule. Each point reports the mean SR over five independent runs with different random seeds, and the error bars indicate one standard deviation.

Fig. 4: Comparison of partial map generation methods. Left: crop along path (PONI); Middle: room visible along path; Right: our VisRC method.

Fig. 5: Physical chair-search sequence at recorded states 57, 77, and 93. Left: RGB images with recorded segmentation overlays (chair in red). Right: accumulated semantic maps over the same map area with full recorded trajectories in blue. The first row overlays unfiltered chair prediction scores on unknown cells: brighter yellow/orange indicates higher scores, while darker purple/black indicates lower scores. The second RGB view shows part of an office area containing two rows of desks and chairs. The highest-response region also contains a chair, distinct from the one subsequently selected for navigation. Green circles mark the chair location obtained from later semantic observations, not surveyed ground truth. The last two rows show target detection and navigation using the previously mapped chair location without prediction overlays. In the final frame, only part of the chair is visible and no chair mask is produced, but its previously mapped location remains available.

Fig. 6: Examples of complete semantic maps generated by composing room maps with matching room categories within floorplan layouts.

Fig. 7: Examples of floorplans with room-category labels under the unified 12-label vocabulary.

Fig. 8: Examples of visibility-constrained partial observations accumulated along VisRC trajectories.

Explore similar work

Multi-Scale Gaussian-Language Map for Zero-shot Embodied Navigation and Reasoning
May 3, 2026Sixian Zhang, Yiyao Wang, Xinhang Song +3Semantic Mapping Vision-Language Navigation
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing · University of Chinese Academy Sciences, Beijing · Institute of Computing Technology, Chinese Academy of Sciences, Beijing
Instance-Enriched Semantic Maps for Visual Language Navigation
Jul 14, 2026Jiho Hong, Eunae Kang, Sanghyun Kim +1Vision-Language Navigation Semantic Mapping
Department of Mechanical Engineering, Kyung Hee University, Yongin-si, 17104, Republic of Korea · Advanced Institutes of Convergence Technology (AICT), Suwon, 16229, Republic of Korea · School of Mechanical Engineering, Kyungpook National University, Daegu, 41566, Republic of Korea
FloVerse: Floor Plan-Guided Multi-Modal Navigation
Jun 12, 2026Weiqi Huang, Shuangyi Dong, Jiaxin Li +3Object Goal Navigation Floor Plan Generation
School of Computer Science & Technology, Beijing Institute of Technology

May 3, 2026, cs.CVJ/K move · Enter open · S save

Multi-Scale Gaussian-Language Map for Zero-shot Embodied Navigation and Reasoning

Sixian Zhang, Yiyao Wang, Xinhang Song, Keming Zhang+2

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing · University of Chinese Academy Sciences, Beijing · Institute of Computing Technology, Chinese Academy of Sciences, Beijing

Semantic Mapping Vision-Language Navigation Point Clouds Navigation

Open paper AskPDF arXiv

Understanding the geometric and semantic structure of environments is essential for embodied navigation and reasoning. Existing semantic mapping methods trade off between explicit geometry and multi-scale semantics, and lack a native interface for large models, thus requiring additional training of feature projection for semantic alignment. To this end, we propose the multi-scale Gaussian-Language Map (GLMap), which introduces three key designs: (1) explicit geometry, (2) multi-scale semantics covering both instance and region concepts, and (3) a dual-modality interface where each semantic unit jointly stores a natural language description and a 3D Gaussian representation. The 3D Gaussians enable compact storage and fast rendering of task-relevant images via Gaussian splatting. To enable efficient incremental construction, we further propose a Gaussian Estimator that analytically derives Gaussian parameters from dense point clouds without gradient-based optimization. Experiments on ObjectNav, InstNav, and SQA tasks show that GLMap effectively enhances target navigation and contextual reasoning, while remaining compatible with large-model-based methods in a zero-shot manner. The code is available at https://github.com/sx-zhang/GLMap.

Abstract

Explore similar work

Multi-Scale Gaussian-Language Map for Zero-shot Embodied Navigation and Reasoning

Instance-Enriched Semantic Maps for Visual Language Navigation

FloVerse: Floor Plan-Guided Multi-Modal Navigation