Organizations: Tsinghua University, Beijing, China · School of Software, Tsinghua University, Beijing, China · Electrical and Computer Engineering, The University of Hong Kong, Hong Kong SAR, China · Qiyuan Lab, Beijing, China
Industrial inspection robots must reach work areas and specific equipment. These destinations span spatial scales from regions to objects, so robots must relate each target to its surroundings to interpret the task and determine where to go. Existing methods support multi-scale navigation through rich environment representations, but at substantial computational and storage cost. These costs limit deployment where wireless coverage is incomplete and robots must rely on onboard computation to keep navigating. We introduce EMS-Nav, a biologically inspired multi-scale navigation framework for online, on-device operation. Its core innovation is to organize perception, spatial understanding, and navigation around regions, making local spaces the basic units of both environmental understanding and computation. EMS-Nav groups objects and their relationships into functionally meaningful spaces and builds a hierarchy from local areas to the broader environment. This organization supports efficient map construction alongside the interpretation and execution of language goals at different scales. In multi-scale navigation experiments, EMS-Nav achieves the highest success rates across all four target categories: objects, rooms, regions, and specific instances. It outperforms state-of-the-art open-vocabulary mapping and navigation methods, improving overall first-attempt success by 9.39 percentage points over the strongest baseline while reducing mapping overhead and language-model input size. Local deployment in real indoor environments further demonstrates its applicability to on-device embodied navigation. By establishing regions as the organizing basis for perception, spatial understanding, and language-guided action, EMS-Nav provides a system that combines multi-scale understanding with computational efficiency to support diverse language goals on resource-constrained robots.
Figures & tables
Figure 1. EMS-Nav uses regions to connect online mapping with multi-scale spatial parsing. Left: posed RGB-D observations provide superpoint geometry and visual features. Center: GSIF-guided relations form coarse regions that constrain instance association and merging. Right: periodic S2R refinement and vision-language parsing build functional regions and their semantic hierarchy for navigation. Spatial parsing leaves the online coarse partition unchanged. MCLIP denotes MobileCLIP-S2. Posed RGB-D observations pass through superpoint feature extraction, Gaussian fields, a relation graph, and online regions to local instance merging. A periodic branch refines the regions and performs multi-scale spatial parsing.
Figure 2. System architecture and spatial organization. (a,b) Perception, spatial parsing, and navigation draw inspiration from local layout perception, spatial integration, and goal-directed control. (c) Nodes and edges within a functional region share a color; translucent outlines indicate structures one level above. (d,e) CG links objects through relations, while HOV-SG groups them within rooms. EMS-Nav connects functional regions to larger spaces for multi-scale goal interpretation. The left side shows biological inspiration and the division of responsibilities among perception, local regions, and semantic parsing. The right side compares region-guided mapping with other forms of spatial organization.
Figure 3. Observation evidence for organizing local spaces. (a) A top-down view and three viewpoints illustrate repeated co-visibility among local objects. (b) GSIF combines object geometry and architectural boundaries to guide relation proposals. (c) Enlarged Gaussian contours and field directions show how shapes and surrounding layout influence candidate connections. These cues complement visual semantics for region formation; co-visibility alone does not determine membership. The left side shows a local space and three observation viewpoints. The center shows the top-down scene layout and field directions. Enlargements on the right show Gaussian shapes and directional cues around objects.
Figure 4. Cumulative navigation success under object, room, region, and instance constraints, together with overall results. Bars show SR@1; plus and cross markers indicate SR@3 and SR@5, respectively. The vertical axis contains a break. EMS-Nav leads across all four goal categories and attempt budgets, showing that one environment representation supports language goals at different spatial scales. Cumulative navigation success rates for EMS-Nav, HOV-SG, DualMap, and CG, shown for four instruction categories and overall.
Figure 5. Perceptual quality and component comparisons for multi-scale navigation. (a,b) Class-agnostic and semantic instance reconstruction. (c) Language-model input/output tokens across online mapping and querying. (d) Navigation with quantized Qwen3-VL-8B-Thinking and Qwen3.7-plus. (e) Region grouping quality measured by ARI and B 3 F1. Token curves follow model-call order; one isolated EMS-Nav output spike is median-smoothed for display. Top row, left to right: class-agnostic instance reconstruction accuracy, and semantic instance reconstruction accuracy. Bottom row: language-model input and output tokens, navigation success rates with quantized Qwen3-VL-8B-Thinking and Qwen3.7-plus, and ARI and B-cubed F1 for S2R and alternative region grouping methods. Token curves follow model-call order, with one isolated EMS-Nav output spike median-smoothed for display. CC is distance-based connected components, DBSCAN is density-based clustering, MLP is a multilayer perceptron, and S2R is our region formation method.
Figure 6. Mapping model size and processing cost. (a) Parameter counts for three mapping stages; asterisks indicate no additional learned module. (b) Fitted per-frame mapping-time estimates at original sequence lengths of 2k, 3k, and 4k frames. Dashed double-headed arrows show processing-time ratios between adjacent methods; the vertical axis is logarithmic. EMS-Nav maintains low mapping costs with modest growth over the tested range. Left: model parameter counts for three mapping stages. Right: per-frame mapping times and relative processing-time ratios for four methods at three input sizes.
Figure 7. Multi-scale functional regions in HM3D scenes qyAac8rV8Zk (left) and mL8ThkuaVTM (right). Each example shows the original scene, its large regions, and functional regions within an open space; lines link corresponding local views. Living, dining, cooking, and other functions remain distinguishable within the larger space, illustrating spatial organization beyond room boundaries. Two scenes show functional regions within local open spaces, alongside the original three-dimensional observations and floor-level regions. Red lines link corresponding local views.
Figure 8. Real-world deployment and resource measurements on an RTX 4060 Ti. (a) Representative indoor environments, the RGB-D device, and collection routes illustrate the observation settings. (b) GPU utilization and memory usage during operation show the computational workload and memory demand of local processing. These measurements complement the navigation results reported in the text when assessing on-device operation. Deployment environments are shown alongside GPU utilization and memory-usage traces.
Figure 9. Ablation of functional regions and multi-scale organization with a shared instance map. Bars show SR@1 on the left axis; plus and cross markers show SR@3 and SR@5 on the right axis. Blue dashed lines mark original HOV-SG SR@1; blue and ochre arrows show differences from this reference. The four settings separate the contributions of functional regions and higher-level organization. Labels are rounded to one decimal place. A closed-frame plot with two vertical axes contains all SR@1 bars, SR@3 plus signs, SR@5 cross markers, blue HOV-SG reference lines, and ten groups of difference arrows.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
EMS-Nav
DualMap
HOV-SG
CG
Constraint
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
Object
55.68
69.76
75.41
52.23
65.86
70.68
51.03
67.57
73.12
30.75
38.63
40.48
Room
44.63
60.42
65.95
34.64
50.50
58.61
39.21
54.96
61.49
18.39
25.03
26.87
Region
39.32
56.51
64.12
27.37
44.98
52.62
23.69
36.94
46.26
9.10
12.93
14.64
Instance
43.59
63.03
71.77
34.07
50.89
57.71
30.61
44.51
51.13
12.00
16.33
17.60
Overall
44.51
61.56
68.70
35.12
51.43
58.50
33.93
48.51
55.72
15.68
20.98
22.61
Appendix
Table 1. Cumulative navigation success rates (%) by constraint type and overall.
Flat instances
HOV-SG-style hierarchy
HOV-SG-style + regions
EMS-Nav + regions
Constraint
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
SR@1
SR@3
SR@5
Object
32.12
44.74
47.88
33.50
46.13
49.05
53.21
68.51
74.72
55.68
69.76
75.41
Room
17.31
25.20
27.61
24.04
32.06
34.29
40.41
55.24
64.06
44.63
60.42
65.95
Region
8.01
10.34
10.83
17.73
23.55
24.47
37.91
56.22
64.02
39.32
56.51
64.12
Instance
5.92
7.62
7.84
14.61
20.49
21.32
40.23
57.91
66.31
43.59
63.03
71.77
Overall
13.42
18.53
19.83
20.78
28.22
29.76
41.63
58.42
66.40
44.51
61.56
68.70
Appendix
Table 2. Complete navigation results (%) for four spatial organization settings, all using EMS-Nav instance representations.
Language model
SR@1
SR@3
SR@5
Quantized Qwen3-VL-8B-Thinking
44.51
61.56
68.70
Qwen3.7-plus
45.27
62.02
68.86
Appendix
Table 3. Cumulative navigation success rates (%) with different language models.
Class-agnostic
Semantic
Method
AP
AP50
AP25
AP
AP50
AP25
EMS-Nav
13.360
26.010
45.230
7.626
12.652
19.727
CG
13.630
29.700
43.530
7.150
12.740
17.340
DualMap
11.840
22.030
28.150
8.170
13.600
16.920
HOV-SG
18.230
29.690
42.350
10.530
15.040
21.000
EMS-Nav w/o merge
12.350
24.896
45.103
7.196
12.365
19.616
Appendix
Table 4. Complete AP results for instance mapping.
Embodied navigation requires agents to translate heterogeneous goals and visual observations into actions across tasks, environments, and robot embodiments. Modern vision-language models (VLMs) already encode spatial priors for visual grounding, spatial reasoning, and pointing, but these capabilities are rarely elicited directly for robot control. Existing navigation systems instead rely on task- or embodiment-specific components, fragmenting perception, reasoning, and action while offering limited generalization. Here we present LightNav-0, a compact generalist embodied navigation model that elicits the spatial intelligence of a pretrained VLM and aligns it with navigation, without task-specific prediction heads. LightNav-0 represents diverse navigation tasks through a unified token interface: dual-channel pointing expresses task-, scene-, and embodiment-agnostic spatial intent, while a residual vector-quantized action tokenizer maps this intent to precise, embodiment-specific trajectories. Together with temporally aware visual history compression, ER mid-training, supervised fine-tuning, and reinforcement learning, this formulation supports instruction following, open-vocabulary object navigation, and visual tracking within a single model. The navigation training corpus spans 2K+ scenes and 4K+ hours of embodied navigation data. LightNav-ER, the embodied-reasoning checkpoint used to initialize LightNav-0, attains the highest complete-set average across 8 embodied-reasoning benchmarks, while LightNav-0 achieves state-of-the-art monocular success rates across all 10 public navigation simulation settings. Real-world evaluations further demonstrate zero-shot generalization across robot embodiments, diverse scenes, and static and dynamic targets. These results establish compact VLMs as a unified and transferable backbone for generalist embodied navigation.
Deploying navigation systems at scale requires a recipe that minimizes sensor assumptions, generalizes across robot embodiments, and trains efficiently. Yet, today's best systems depend on depth sensors, multi-camera rigs, or pre-built maps, limiting the hardware they support and increasing deployment cost. We introduce Robostral Navigate, an 8B vision-language model built around this scalability objective. The model consumes only a stream of monocular RGB images - the most ubiquitous sensor across robotic platforms and predicts waypoints by pointing to the next target location in the current camera view. Operating purely in image space, rather than robot-specific coordinates, makes the policy naturally robust to changes in camera intrinsics and scene scale, enabling deployment across wheeled, legged, and aerial robots without recalibration. We generate 2.4 million trajectories across 350k simulated scenes to reduce the reliance on real-world data collection and scale easily. We further introduce a prefix-caching training recipe that packs entire episodes into single training sequences, reducing training tokens by 22x and cutting training time from months to days. A tree-based attention mask prevents conditioning on previous ground-truth actions, encouraging visually grounded action prediction, and reinforcement learning is used to further improve exploration and recovery capabilities. On the Room-to-Room and Room-Across-Room in Continuous Environments (R2R-CE and RxR-CE) benchmarks, Robostral Navigate sets a new state of the art. On R2R-CE, it achieves a 77.4% success rate, surpassing the best monocular method by 10.5 points and the strongest depth- or multi-camera system by 5.3 points despite using only a single RGB camera. On RxR-CE, it reaches 75.1% success rate, outperforming all monocular baselines.
Object navigation requires an agent to locate a target in an unknown environment through visual observations. Existing methods typically rely on open-vocabulary detectors or vision-language models (VLMs) to answer where to search, but often overlook what not to trust - which semantic cues are unreliable. Open-vocabulary perception is prone to systematic misleading evidence: false positives, outdated static priors, and repeated failed exploration due to lack of embodied verification, which contaminates mapping and decision-making. Such errors are rooted in structured object relations in real-world scenes. To address this, we propose DB-Nav, a framework that reshapes the search space via dual relational biases. It factorizes target-centric relations into an Activation Bias (propagates contextual evidence) and an Inhibition Bias (suppresses unreliable regions via perceptual confusion and action-level falsification). These biases are unified into a Relational Activation-Inhibition Exploration Graph that modulates frontier exploration values using online observations and failed accesses. Experiments on ObjectNav benchmarks show that DB-Nav significantly outperforms existing methods in success rate (SR) and Success weighted by Path Length (SPL), offering a lightweight, interpretable, and robust navigation framework without costly online VLM reasoning.
Weitao An, Chenghao Xu, Xu Yang +1
School of Electronic Engineering, Xidian University, Xi’an 710071, China · School of Information Science and Engineering, Hohai University, Nanjing 210098, China