Active Spatial Inspection for Effective and Efficient Embodied Exploration

Wenbin Wang; Xiang Bai; Yizhao Wang; Hang Sun; Dong Ren; Jie Qin; Qingquan Li; Bing Wang

cs.ROSep 18, 2026

Active Spatial Inspection for Effective and Efficient Embodied Exploration

Authors: Wenbin Wang, Xiang Bai, Yizhao Wang, Hang Sun, Dong Ren, Jie Qin, Qingquan Li, Bing Wang

Organizations: The Hong Kong Polytechnic University · China Three Gorges University · Huazhong University of Science and Technology · Nanjing University of Aeronautics and Astronautics ·

Abstract

Achieving high task success and efficiency remains a central pursuit in embodied exploration. Existing frameworks typically guide agent behavior through a spatially coarse and indirect assessment of suggestive cues and directions, yet such designs may struggle to judge cue sufficiency and the need for further inspection, leading to early termination or excessive continuation and ultimately reducing task success and efficiency. This work rethinks embodied exploration from a spatially explicit standpoint and introduces the spatial-inspection-guided ACtive Exploration (ACE) framework coupling evidence-grounded perception with exposure-informed movement and establishing a spatially resolved decision paradigm. Evidence-grounded perception strengthens inferential validity by integrating granular localization with focused verification, turning suggestive cues into precise decisive visual support. Exposure-informed movement promotes directional judgment through prospective prioritization and retrospective suppression, selecting promising directions for efficient spatial progress. ACE alleviates the tension between preventing early termination and avoiding unnecessary continuation. Extensive experiments demonstrate that ACE achieves 18.0% higher navigation task success and 10.3% higher exploration efficiency for question answering than prior state-of-the-art baselines, advancing effective and efficient embodied exploration.

Figures & tables

Fig. 1: Illustration of the two spatial assessment bottlenecks in embodied exploration and how ACE addresses them. (A) Existing methods often reason over coarse visual units, leading to ambiguous termination decisions. ACE instead grounds task-relevant cues into explicit spatial candidates and verifies the localized content to establish reliable evidence for termination. (B) Existing methods indirectly indicate whether a space has already been inspected. Different frontiers may lead to substantially overlapping observations and unnecessary movement. ACE explicitly relates each frontier to the space it exposes, suppressing exhausted directions while favoring those that can reveal useful information. Together, the two aspects illustrate the shift from coarse and indirect assessment to spatially explicit evidence assessment and frontier evaluation.

Fig. 2: Overview of the proposed ACE framework. Given the current observations and accumulated knowledge, (A) Discriminative Probing localizes and verifies evidence for termination. (B) Cross-Temporal Frontier Association links transient frontier instances across decision steps, providing a persistent association basis for frontier evaluation and regulation. If decisive evidence is unavailable, (C) Prospective Prioritization estimates frontier utility from visible task cues and hidden-space potential. (D) Retrospective Suppression rejects exhausted directions and penalizes low-progress exploration using accumulated feedback. The remaining frontier with the highest regulated utility is selected for continued exploration.

Baselines
Method	LLM-Match (%) $\uparrow$	LLM-Match SPL (%) $\uparrow$
GPT-4o [ 1 ]	35.9	–
Multi-Frame* [ 6 ]	41.8	7.5
VLM-based exploration
Explore-EQA [ 16 ]	46.9	23.4
CG + Frontier [ 1 , 2 ]	47.2	33.3

TABLE I: Performance comparison on A-EQA 184-subset. The best and second-best results are highlighted in bold and underlined, respectively.

Supervised methods
Method	SR (%) $\uparrow$	SPL (%) $\uparrow$
BC [ 71 ]	5.4	1.9
DAgRL [ 72 ]	18.3	7.9
DAgRL+OD [ 66 ]	37.1	19.8
Uni-NaVid [ 73 ]	39.5	19.8
OVSegDT [ 74 ]	40.1	20.9

TABLE II: Performance comparison on HM3D-OVON.

Method	Full set		278-subset
	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	SPL $\uparrow$
Supervised methods
SenseAct-NN Skill Chain [ 11 ]	29.5	11.3	–	–
SenseAct-NN Monolitic [ 11 ]	12.3	6.8	–	–
MTU3D [ 36 ]	47.2	27.7	–	–
AstraNav-Memory [ 51 ]	62.7	56.9	–	–

TABLE III: Performance comparison on the GOAT-Bench.

Method	DP	PP (AHS)	RS		LLM-Match $\uparrow$	$\Delta$	LLM-Match SPL $\uparrow$	$\Delta$
			VFR	HLP
Baseline-1 (holistic)	–	–	$\times$	$\times$	51.4	–	35.1	–
Baseline-2 (separated)	$\times$	$\times$	$\times$	$\times$	51.4	+0.0	37.2	+2.1
DP w/o verify		$\times$	$\times$	$\times$	53.1	+1.7	39.9	+2.7
DP	$\checkmark$	$\times$	$\times$	$\times$	56.5	+3.4	41.8	+1.9
DP +AHS	$\checkmark$	$\checkmark$	$\times$	$\times$	57.6	+1.1	41.1	-0.7

TABLE IV: Ablation study of ACE main components on the A-EQA 41-subset. Discriminative Probing (DP) refers to region-based granular localization and focused verification. Prospective Prioritization (PP) refers to the Active Hidden-Space (AHS) score. Retrospective Suppression (RS) includes Visual Frontier Rejection (VFR) and History-aware Low-progress Penalization (HLP). For incremental ablations and baseline comparisons,

\Delta

is computed relative to the adjacent previous row; for single-component removal ablations,

\Delta

is computed relative to the full ACE framework.

Method	DP (EVPLV)	PP (AHS)	RS		SR $\uparrow$	$\Delta$	SPL $\uparrow$	$\Delta$
			VFR	HLP
Baseline-1 (holistic)	–	–	$\times$	$\times$	68.8	–	41.7	–
Baseline-2 (separated)	$\times$	$\times$	$\times$	$\times$	69.3	+0.5	43.2	+1.5
EVPLV w/o verify		$\times$	$\times$	$\times$	70.5	+1.2	44.1	+0.9
EVPLV	$\checkmark$	$\times$	$\times$	$\times$	72.9	+2.4	45.4	+1.3
EVPLV+AHS	$\checkmark$	$\checkmark$	$\times$	$\times$	75.7	+2.8	47.6	+2.2

TABLE V: Ablation study of ACE main components on the GOAT-Bench 278-subset. Discriminative Probing (DP) refers to Evidential Visual Point Localization and Verification (EVPLV).

Dataset	VLM	Cost	LLM-Match/SR $\uparrow$	SPL $\uparrow$
A-EQA	Qwen3-VL-32B	$0.16/$ 0.65	50.5	37.7
	Qwen3.5-27B	$0.08/$ 0.67	54.6	45.8
	Qwen3.5-35B	$0.06/$ 0.44	53.6	49.7
	Qwen3.5-397B	$0.17/$ 1.00	56.9	53.4
	GPT-4o	$5.00/$ 15.00	60.4	50.2
GOAT-Bench	Qwen3.5-397B	$0.17/$ 1.00	72.4	51.1

TABLE VI: Replacement of VLMs. Cost denotes input/output API price per million tokens.

Fig. 3: Efficiency–quality statistics of regional evidence probing on A-EQA. (a) Average steps to final answer commitment and average GPT-based answer score for “w/o DP” and ACE. (b) Score-conditioned distributions of commitment steps. Each violin represents the step distribution within an answer-score group. The horizontal bar indicates the median step count, and the vertical bar indicates the interquartile range.

Fig. 4: Qualitative comparison of answer commitment. Each example shows the explored trajectory ( ACE vs. w/o DP ), the snapshot used for final answering, the corresponding predicted answer, and step count. The blue box denotes the selected region from the $3\times 3$ regional partition, which is identified by the VLM as the evidence-bearing region for answering.

Fig. 5: Analysis of EVPs. (a) The full ACE and w/o EVPLV are compared when no verified object evidence is available. (b) The EVP candidates are progressively filtered.

Method	SR $\uparrow$	SPL $\uparrow$
ACE w. bbox	74.2	47.7
ACE	76.3	50.5

TABLE VII: Comparison between bbox-based localization and EVP localization on the GOAT-Bench 278-subset.

Fig. 6: Qualitative examples of terminal EVP verification on GOAT-Bench. The orange point and box denote the initially proposed, unverified EVP from the current observation, while the blue point and box denote the close-range verified EVP used for termination. The red box in the bottom-left case marks the challenging hanger target.

Fig. 7: Behavior-level analyses. (a,b) Weak-RTC commitment at

K=5

and varying commitment horizons. (c,d) Exhausted-region reselection at

K=2

and varying minimum revisit gaps. (e,f) Trajectory progress ratio at

K=4

and varying rolling frontier-move windows. Error bars and shaded bands denote 95% episode-level bootstrap confidence intervals.

Panel of Figure 15 — Fig. 8: Qualitative examples of prospective prioritization. Each case lists frontier candidates of a certain step with their RTC and AHS responses, numerical scores, and the resulting frontier utility. The selected frontier is highlighted in purple.

Fig. 9: Comparison of the chosen frontier of each step between ACE and w/o RS .

Panel of Figure 17 — Fig. 10: Successful and failed trajectories on GOAT-Bench.

Embodied AI

Embodied AI studies agents that perceive, reason, and act through bodies in physical or simulated environments.

See all 255 papers

ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop
May 18, 2026Yining Hong, Jiageng Liu, Han Yin +5Active Perception Spatial Reasoning Benchmarks
1Stanford University · 2UCLA · 3Northwestern University
Explore Like Humans: Autonomous Exploration with Online SG-Memo Construction for Embodied Agents
Apr 21, 2026Xu Chen, Shichao Xie, Zhining Gu +7Embodied Navigation Semantic Mapping
Amap, Alibaba Group, China.
FUSE: Active Functional Affordance Grounding through Adaptive Semantic-Geometric Evidence Acquisition
Aug 13, 2026Zhou Chen, Sathyanarayanan N. AakurNext-Best-View Planning Active Sensing
CSSE Department Auburn University, Auburn, AL, 36849

May 18, 2026J/K move · Enter open · S save

ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop

Yining Hong, Jiageng Liu, Han Yin, Manling Li+4

1Stanford University · 2UCLA · 3Northwestern University

Active Perception Spatial Reasoning Benchmarks 3D Spatial Reasoning Embodied Spatial Reasoning AI Agent Benchmarks Active Sensing+2

Open paper AskPDF arXiv

Spatial intelligence unfolds through a perception-action loop: agents act to acquire observations, and reason about how observations vary as a function of action. Rather than passively processing what is seen, they actively uncover what is unseen - occluded structure, dynamics, containment, and functionality that cannot be resolved from passive sensing alone. We move beyond prior formulations of spatial intelligence that assume oracle observations by recasting the observer as an actor. We introduce ESI-BENCH, a comprehensive benchmark for embodied spatial intelligence spanning 10 task categories and 29 subcategories built on OmniGibson, grounded in Spelke's core knowledge systems. Agents must decide what abilities to deploy - perception, locomotion, and manipulation - and how to sequence them to actively accumulate task-relevant evidence. We conduct extensive experiments on state-of-the-art MLLMs and find that active exploration substantially outperforms passive counterparts, with agents spontaneously discovering emergent spatial strategies without explicit instructions, while random multi-view often adds noise rather than signal despite consuming far more images. Most failures stem not from weak perception but from action blindness: poor action choices lead to poor observations, which in turn drive cascading errors. While explicit 3D grounding stabilizes reasoning on depth-sensitive tasks, imperfect 3D representation proves more harmful than 2D baselines by distorting spatial relations. Human studies further reveal that unlike humans who seek falsifying viewpoints and revise beliefs under contradiction, models commit prematurely with high confidence regardless of evidence quality, exposing a metacognitive gap that neither better perception nor more embodied interaction alone can close.

Abstract

Embodied AI

Explore similar work

ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop

Explore Like Humans: Autonomous Exploration with Online SG-Memo Construction for Embodied Agents

FUSE: Active Functional Affordance Grounding through Adaptive Semantic-Geometric Evidence Acquisition