Object-goal navigation (ObjectNav) requires an embodied agent to find an instance of a specified category in an unseen environment and stop near it. The agent typically relies on an object detector but must judge online from a limited egocentric video stream, sometimes a single frame, whether a detected candidate is the target. Detectors often confidently mistake similar objects, such as a sofa for a bed, and stopping at a false positive fails the episode. Prevailing systems also approach candidates greedily, leaving them occluded or partially observed. We present CLOSER, an evidence-aware active verification framework that treats stopping as closed-loop evidence acquisition, coupling vision-language model (VLM) verification with active perception. CLOSER fine-tunes an open-source VLM to label each view of a candidate tracked across frames as target, non-target, or uncertain, selects complementary views with an embedding-based diversity gate, and fuses their per-view probabilities with a lightweight network. If the fused judgment remains uncertain, a diffusion policy drives the agent toward the candidate using only the tracked mask and depth until the decision becomes decisive. CLOSER runs locally and plugs into existing ObjectNav systems without retraining them. On Gibson and Habitat-Matterport 3D, it raises average success rate by 3.6 points and success weighted by path length by 2.0 points, showing that verifying evidence before committing improves embodied navigation.
Figures & tables
Fig. 1: With CLOSER, the agent approaches an uncertain candidate and decides from closer views. Given the goal bed , the view at (1) shows a distant, partially occluded candidate and does not suffice for a decision. CLOSER therefore labels the candidate uncertain , and the agent approaches it to gather additional views (orange, dotted). At (2), the closer view reveals a sofa confuser; CLOSER labels it no , and the agent resumes exploration (blue, solid). At (3), the view of the next candidate suffices; CLOSER labels it yes , and the agent moves to the bed and stops near it. The bottom panels show the agent’s view and CLOSER’s label at each numbered location.
Fig. 2: Closed-loop inference and training of CLOSER. (A) Grounded-SAM, ByteTrack, and SAM track the candidate and refresh its mask in each RGB-D frame. (B) The evidence cache retains views selected by a CLIP-based diversity gate. (C) Per-view tri-state probabilities from the single-view perceiver are fused by a lightweight MLP. (D) After an uncertain decision, a Diffusion Transformer generates waypoints from the mask and depth. Frames observed along the waypoints re-enter panels A and B (uncertain loop). The lower panels depict the perceiver training pipeline ( Verifier Training ) and the construction of expert waypoint chunks ( Policy Training ).
Sufficient Evidence
Insufficient Evidence
Target candidate
yes
uncertain
Non-target candidate
no
uncertain
TABLE I: Evidence-aware label assignment.
Setting
Value
Target-overlap thresholds (η+,η−)
(0.25,0.05)
Minimum mask-to-image ratio α
0.02
Minimum box-fill ratio β
0.40
Maximum mean mask depth γ
3.0 m
Cache capacity K
4 views
View-diversity threshold δview
0.18
TABLE II: Implementation settings for CLOSER.
Method
OV
Gibson
HM3D v1
HM3D v2
SR(%) ↑
SPL(%) ↑
SR(%) ↑
SPL(%) ↑
SR(%) ↑
SPL(%) ↑
Plug-in comparison
ZSON [ 54 ]
✓
31.3
12.0
25.5
12.6
28.7
13.1
+ CLOSER
✓
35.6 (+4.3)
14.5 (+2.5)
29.8 (+4.3)
15.1 (+2.5)
32.5 (+3.8)
14.7 (+1.6)
DD-PPO [ 55 ]
×
34.8
14.0
27.9
14.2
33.5
15.6
+ CLOSER
×
38.9 (+4.1)
16.3 (+2.3)
32.1 (+4.2)
15.6 (+1.4)
36.9 (+3.4)
17.8 (+2.2)
TABLE III: End-to-end ObjectNav results. SR/SPL of eight ObjectNav systems and their CLOSER plug-in variants on Gibson, HM3D v1, and HM3D v2 (OV: open-vocabulary). Each shaded “+ CLOSER” row reports the absolute gain over its baseline in parentheses; CLOSER results are means over three seeds (standard deviations ≤ 0.9 SR and ≤ 0.6 SPL points). CLOSER improves both metrics for every system on every dataset.
Fig. 3: Paired ObjectNav gains from CLOSER. (a) Absolute SR gains grouped by dataset and by the native candidate-verification mechanism (shading, left to right: none, map-level heuristic, and candidate-level fusion). (b) Changes in SR and SPL for the same 24 system–dataset pairs; each arrow starts at the baseline and ends at its “+ CLOSER” result. Data from Table III .
Fig. 4: Closed-loop verification in one successful episode with target category sofa . (a) Evidence timeline of the two candidate events (S: step index). In event A, the initial view is deferred as uncertain , and the view retained at step 37 reveals the candidate to be a bed, so the agent discards the candidate and resumes exploration. In event B, the fused decision remains uncertain at step 72 and becomes yes at step 75, so the agent stops. The quoted text shows the perceiver rationales for the two views of event A, decoded here for illustration. (b) Top-down path and semantic map of the episode; the colored path segments correspond to the timeline arrows in (a).
Fig. 5: Real-world verification on a quadruped robot. The verification stack runs on the onboard camera stream of a Unitree Go2 in two indoor scenes (one scene per row, frames ordered left to right; insets show the onboard view). Stacked blocks occlude the candidate (a chair, top; an armchair, bottom), so the initial views are deferred as uncertain . As the robot moves around the blocks, clearer views enter the cache, and the fused decision becomes decisive once an unobstructed view is retained (rightmost insets).
Variant
SR ↑
SPL ↑
FS ↓
FR ↓
FB ↓
VLFM native stop
63.6
32.5
19.4
13.8
–
Binary GPT-4o-mini (API)
64.9
33.3
11.2
12.1
–
Binary local
64.4
33.0
13.6
14.2
–
Tri-state + diverse cache (static)
65.3
33.6
8.7
15.1
4.2
+ Diffusion policy (CLOSER)
66.1
34.1
6.4
9.3
9.8
TABLE IV: Closed-loop ablation of the verification layer. All variants are built on VLFM and evaluated on the HM3D v2 validation episodes. FS / FR / FB: episode-level false stop, false reject, and fallback trigger (%).
Fig. 6: Episode outcomes of the closed-loop ablation. Each episode is assigned to exactly one outcome; the right column shows the total stop-attributable error (FS + FR). Data from Table IV .
Verifier / variant
Clear → yes ↑
Low-Q tgt → unc. ↑
Clear → no ↑
Low-Q cnf → unc. ↑
Qwen2.5-VL zero-shot
68.7
42.3
55.1
39.8
Binary SFT ( yes / no )
87.4
0.0
78.6
0.0
Binary SFT + threshold
79.8
82.1
69.4
80.6
Direct-label SFT, single view
83.6
67.4
73.2
65.8
+ Rendered regimes + rationale
85.9 (+2.3)
78.7 (+11.3)
77.1 (+3.9)
77.3 (+11.5)
Full CLOSER stack
89.1 (+3.2)
85.4 (+6.7)
84.6 (+7.5)
83.8 (+6.5)
TABLE V: Verifier comparison on 5,184 held-out Grounded-SAM detections. Clear columns: correct semantic decisions on clear observations; Low-Q columns: correct deferral of insufficient observations (zero for the pure binary verifier, which cannot defer). The lower block reports the CLOSER training stages, the first two on single views; parenthesized values give the gain over the previous stage. All values in %.
Evaluation domain
Binary SFT + threshold
CLOSER (tri-state)
Defer ↑
FS ↓
FR ↓
Defer ↑
FS ↓
FR ↓
HM3D held-out (in-domain)
81.4
12.6
11.2
84.6
7.5
8.4
MP3D candidates (cross-dataset)
63.7
17.9
15.8
82.1
8.6
9.3
Rendered hard regimes
46.8
23.4
20.7
79.2
9.8
10.9
TABLE VI: Deferral quality under distribution shift. The deferral threshold of the binary verifier is selected once on the HM3D validation split and kept fixed. Defer: fraction of insufficient-evidence candidates correctly deferred; FS / FR: candidate-level false stop and false reject. All values in %.
Fig. 7: Deferral under distribution shift. The fixed binary threshold and CLOSER are evaluated on held-out HM3D candidates, MP3D candidates, and the rendered hard regimes of Table I . Annotations show changes from the in-domain to the rendered setting; the shaded region shows the gap between the methods. Data from Table VI .
Evidence input
Acc. ↑
Low-Q ↑
FS ↓
Latest frame only
80.7
78.9
11.4
Latest- K frames
83.5
81.2
9.2
CLIP-diverse (ours)
86.1
84.6
7.5
TABLE VII: Ablation of cache construction. The learned components are fixed; multi-view rows use K=4 . Acc.: tri-state accuracy; Low-Q: fraction of insufficient-evidence candidates correctly deferred; FS: candidate-level false stop. All values in %.
Policy
Correct ↑
Confuser ↑
Path (m) ↓
FB ↓
In-place look-around
32.4
26.1
0.00
0.0
Straight-line policy
61.7
54.3
2.84
37.6
Viewpoint re-verif. (STRIVE) [ 11 ]
70.2
63.8
2.21
25.4
Voronoi–Bézier det. BC
76.5
70.9
2.08
18.1
Voronoi–Bézier diffusion
83.6
78.4
1.69
12.7
TABLE VIII: Approach-policy comparison on initially uncertain candidates. Correct: correct yes / no decision before the close-range fallback; Confuser: rejection rate on initially uncertain confusers; FB: fallback trigger rate on this candidate subset. All values except Path are in %.
Fig. 8: Resolution accuracy, extra path, and fallback rate of approach policies on the initially uncertain candidate subset. Data from Table VIII .
Verifier
Local
Views
Tri-state accuracy ↑
Latency (s) ↓
Memory (GB) ↓
Cost/1k queries ↓
GPT-4o-mini API (tri-state)
No
1
82.4
1.38
–
0.41
Qwen2.5-VL zero-shot
Yes
1
52.6
0.31
16.9
0
CLOSER single-view perceiver
Yes
1
80.7
0.28
17.2
0
CLOSER perceiver + raw-probability MLP
Yes
≤4
86.1
0.28/view
17.2
0
TABLE IX: Verifier runtime and deployment cost. Local VLM latency and memory are measured on a single NVIDIA A100 40 GB GPU; API cost is in USD per 1,000 queries. Latency of the perceiver–MLP stack is per retained view; each MLP fusion adds 0.1 ms.
Confusion
Detector only
Binary verifier
CLOSER
Sofa as bed
79.2
53.1
35.4
Stool as chair
70.6
45.3
28.8
Sink as toilet
64.8
42.7
27.1
Other confusions
46.3
30.2
19.6
TABLE X: Candidate-level false-stop rate (%) on characteristic confuser pairs (VLFM+CLOSER, HM3D v2). “A as B” denotes a candidate of category A detected while searching for target category B; “Other confusions” pools the unlisted non-target confusion types.
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
Zero-shot object-goal navigation (ZSON) in open-vocabulary scenarios is challenging, as it requires a robot to locate an arbitrarily specified object in an unseen environment without task-specific training. Currently, the task still suffers from high latency and limited accuracy due to redundant perception pipelines and insufficient evidence for reliable target confirmation. In this letter, we reframe ZSON as an evidence-driven perception-to-decision problem and present AECNav, a training-free pipeline built on three components: i) Evidence-gated perception, which utilizes a shared encoding across all reasoning stages to establish a unified semantic basis and eliminate redundant computations; ii) Evidence consolidation, which aggregates detections into cluster-level log-odds beliefs. This explicitly separates genuine target support from the false confidence of visually similar distractors, while treating the absence of expected detections as negative evidence; and iii) Active evidence acquisition, which sustains productive exploration under weak semantic cues by selecting frontiers that maximize information gain at minimal traversal cost. As a result, AECNav significantly outperforms previous methods and achieves state-of-the-art success rates of 84.7%, 57.3%, and 51.3% on HM3D-v2, HM3D-OVON, and MP3D, respectively, with substantially lower inference overhead, and attains 95% success across 40 trials on a physical quadruped robot at roughly 5Hz. Code will be made publicly available upon acceptance.
Guanlin Liu, Shaobin Ling, Renyuan Liu +2
The Chinese University of Hong Kong, Shenzhen · The Hong Kong University of Science and Technology (Guangzhou)
Pretrained VLMs bring rich world knowledge and semantic priors to object-goal navigation, and fine-tuning a VLM into a navigation policy has become a common recipe. Despite this progress, success rates remain low, and it is unclear which capability these fine-tuned policies lack. Our diagnosis identifies goal grounding as the dominant failure. These policies explore well, yet often fail to notice that the goal is already in view or to estimate how far away it is. To address this failure, we ground the goal explicitly. Explicit grounding substantially improves success and reaches state-of-the-art results on six benchmarks across three scene collections. We also correct annotation defects in these benchmarks, so that future methods are evaluated more faithfully. Our analysis on the clean benchmarks shows that, with ground-truth grounding, our policy nearly matches the expert that generated its demonstrations, leaving the headroom in perception rather than policy learning. Our work points to goal grounding as a missing piece of object-goal navigation and offers a simple and effective solution to the task.