Does Peer Observation Help? Vision-Sharing Collaboration for Vision-Language Navigation
Authors: Qunchao Jin, Yiliao Song, Qi Wu
Organizations: Australian Institute for Machine Learning, Adelaide University · Responsible AI Research Centre, Australian Institute for Machine Learning
Vision-Language Navigation (VLN) systems are fundamentally constrained by partial observability, as an agent can only accumulate knowledge from locations it has personally visited. As multiple robots increasingly coexist in shared environments, a natural question arises: can agents navigating the same space benefit from each other's observations? In this work, we introduce Co-VLN, a minimalist, model-agnostic framework for systematically investigating whether and how peer observations from concurrently navigating agents can benefit VLN. When independently navigating agents identify common traversed locations, they exchange structured perceptual memory, effectively expanding each agent's receptive field at no additional exploration cost. We validate our framework on the R2R benchmark under two representative paradigms (the learning-based DUET and the zero-shot MapGPT), and conduct extensive analytical experiments to systematically reveal the underlying dynamics of peer observation sharing in VLN. Results demonstrate that vision-sharing enabled model yields substantial performance improvements across both paradigms, establishing a strong foundation for future research in collaborative embodied navigation.
Figures & tables
Figure 1: Motivation of peer observation sharing in VLN. (a) A single agent navigates independently, with its perception limited to personally visited regions while large portions of the environment remain unobserved, resulting in only 35% spatial coverage. (b) When a peer agent independently navigates the same environment, their trajectories may overlap at a spatial section (yellow star), enabling knowledge exchange that expands perceptual coverage from 35% to 65% without additional exploration.
Figure 2: Overview of the Co-VLN framework. Given multiple agents navigating in a shared environment, Co-VLN operates in three stages: (1) each agent independently explores and builds its own topological graph; (2) spatial overlap between the two graphs is detected via embedding-based matching or viewpoint ID matching; (3) the graphs are merged through anchor nodes identified in Stage 2, producing an enriched navigation graph that expands each agent’s knowledge beyond its own exploration.
Method Paradigm
Methods
TL
NE ↓
OSR ↑
SR ↑
SPL ↑
PREVALENT [ 18 ]
10.19
4.71
–
58
53
VLN ↻ BERT [ 20 ]
12.01
3.93
69
63
57
HAMT [ 10 ]
11.46
3.62
73
66
61
Supervised
HOP+ [ 36 ]
11.76
3.49
–
67
61
Learning
TD-STP [ 59 ]
–
3.22
76
70
63
BEVBert [ 1 ]
14.55
2.81
84
75
64
Table 1: Comparison with state-of-the-art methods on the R2R val unseen. ↑ : higher is better, ↓ : lower is better. MapGPT ‡ uses Qwen3-VL-32B as the MLLM. † denotes the use of oracle viewpoint IDs for spatial overlap detection. ∗ denotes applying CLIP ViT-B16 as image features. Best results are bold .
Methods
TL
NE ↓
OSR ↑
SR ↑
SPL ↑
Seq2Seq [ 25 ]
8.64
7.37
40
32
30
HPN [ 23 ]
7.62
6.31
40
36
34
CM2 [ 16 ]
11.54
7.02
42
34
28
WS-MGMAP [ 9 ]
10.00
6.28
48
39
34
CWP-CMA [ 19 ]
10.90
6.20
52
41
36
CWP-RecBERT [ 19 ]
12.23
5.74
53
44
39
Table 2: Comparison with state-of-the-art methods on the R2R-CE val unseen. ↑ : higher is better, ↓ : lower is better. Best results are bold .
Figure 5
MLLM
Vision- Sharing
NE ↓
OSR ↑
SR ↑
SPL ↑
InternVL3.5-38B
✗
6.60
50.5
37.0
31.1
✓
6.40 (-0.20)
51.5 (+1.0)
38.0 (+1.0)
31.6 (+0.5)
Gemini-3-Flash †
✗
6.09
40.5
38.0
35.2
✓
5.10 (-0.99)
59.0 (+18.5)
46.0 (+8.0)
40.5 (+5.3)
Qwen3-VL-32B
✗
5.51
59.5
46.5
38.0
✓
4.84 (-0.67)
64.0 (+4.5)
49.5 (+3.0)
41.9 (+3.9)
Table 4: Effect of different MLLMs on Vision-Sharing setting. Best results are bold , second best are underlined . † indicates closed-source models.
Area Group
Vision- Sharing
SR ↑
SPL ↑
OSR ↑
NE ↓
Small ( < 250 m 2 )
✗
56.75
49.95
68.65
3.77
✓
58.73
50.71
72.62
3.62
+ 1.98 ( + 3.49%)
+ 0.76 ( + 1.52%)
+ 3.97 ( + 5.78%)
− 0.15 ( − 3.98%)
Medium (250–450 m 2 )
✗
52.40
43.62
61.74
4.26
✓
55.81
46.75
67.30
3.89
+ 3.41 ( + 6.51%)
+ 3.13 ( + 7.18%)
+ 5.56 ( + 9.01%)
− 0.37 ( − 8.69%)
Table 5: Effect of scene area on Vision-Sharing setting. Shaded rows show absolute improvement with relative improvement (%). Bold indicates the best improvement.
Baseline
Vision- Sharing
Pairing Strategy
NE ↓
OSR ↑
SR ↑
SPL ↑
MapGPT
✗
–
4.80
64.0
52.2
44.7
✓
random pairing
4.66
65.3
53.4
45.7
✓
prior-based pairing
4.36
68.6
55.8
47.3
DUET
✗
–
3.31
80.5
71.5
60.4
✓
random pairing
3.24
81.2
72.6
61.2
✓
prior-based pairing
2.99
82.7
74.5
62.6
Table 6: Effect of trajectory pairing strategies on Vision-Sharing setting. Best results are bold , second best are underlined .
Detection Trigger
Bi- direction
Persistent
SR ↑
SPL ↑
DUET (w/o Vision-Sharing)
71.5
60.4
✓
✓
×
72.4
60.0
✓
×
✓
73.7
61.3
×
✓
✓
74.3
61.5
✓
✓
✓
74.5
62.6
Table 7: Effect of graph fusion strategy. Best results are bold , second best are underlined .
Figure 4: Qualitative comparison . We visualize DUET with and without vision-sharing.