OVAL: Open-Vocabulary Augmented Memory Model for Lifelong Object Goal Navigation

Jiahua Pei; Yi Liu; Guoping Pan; Yuanhao Jiang; Houde Liu; Xueqian Wang

cs.ROApr 14, 2026

OVAL: Open-Vocabulary Augmented Memory Model for Lifelong Object Goal Navigation

Authors: Jiahua Pei, Yi Liu, Guoping Pan, Yuanhao Jiang, Houde Liu, Xueqian Wang

Organizations: Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China

Abstract

Object Goal Navigation (ObjectNav) refers to an agent navigating to an object in an unseen environment, which is an ability often required in the accomplishment of complex tasks. While existing methods demonstrate proficiency in isolated single object navigation, their limitations emerge in the restricted applicability of lifelong memory representations, which ultimately hinders effective navigation toward continual targets over extended periods. To address this problem, we propose OVAL, a novel lifelong open-vocabulary memory framework, which enables efficient and precise execution of long-term navigation in semantically open tasks. Within this framework, we introduce memory descriptors to facilitate structured management of the memory model. Additionally, we propose a novel probability-based exploration strategy, utilizing a multi-value frontier scoring to enhance lifelong exploration efficiency. Extensive experiments demonstrate the efficiency and robustness of the proposed system.

Figures & tables

Fig. 2: Pipeline of our OVAL: The frontier exploration module utilizes depth and pose to build a grid map, followed by frontier selection based on probability map. Then, the open-semantic memory model passes through the autolabel extraction, keyword filter, and similarity calculation to update our proposed memory model. Finally, the navigation module searches for goal synonyms via KMP-queries and navigates waypoints, verifying when to stop upon reaching the goal.

Fig. 3: The workflow of Memory Model Management. The system designs the instances matcher module for objects labeled in memory, calculating similarity and matching to identify new instances (Instance refer to different occurrences of the same object category), thereby preventing interference with previously stored memories.

Method	Open-Vocabulary	Lifelong Memory	HM3D		MP3D
Method	Open-Vocabulary	Lifelong Memory	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$
VLFM [ 7 ]	$\checkmark$	$\times$	53.7	30.6	35.1	16.7
GOAT [ 6 ]	$\times$	$\checkmark$	59.2	31.2	N/A	N/A
OVAL (Ours)	$\checkmark$	$\checkmark$	68.1	33.8	44.1	18.6

TABLE I: Comparison with SOTA methods on HM3D and MP3D datasets (1000 episodes) for lifelong ObjectNav. The best results are in bold . N/A denotes a run failure in the dataset.

Fig. 4: Lifelong ObjectNav Tasks Performance Comparison with GOAT in HM3D. Bar and line charts show SR and SPL comparisons, respectively.

Fig. 5: Ablation study on the performance of varying numbers of lifelong targets in the Lifelong Object Nav Tasks.

Verify “STOP”	Memory Model	Probability Map	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$
$\times$	✓	✓	61.3	31.2
✓	$\times$	✓	56.1	26.0
✓	✓	$\times$	66.8	32.3
✓	✓	✓	68.1	33.8

TABLE II: Ablation study on verify “STOP”, memory model, and probability map exploration strategy method on HM3D datasets.

Method	Open-Vocabulary	Training Free	Lifelong Memory	HM3D		MP3D
Method	Open-Vocabulary	Training Free	Lifelong Memory	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$
PONI [ 37 ]	$\times$	$\times$	$\times$	-	-	31.8	12.1
SemEXP [ 12 ]	$\times$	$\times$	$\times$	-	-	36.0	14.4
L3MVN [ 18 ]	$\times$	$\checkmark$	$\times$	50.4	23.1	-	-
Habitat-Web [ 10 ]	$\checkmark$	$\times$	$\times$	41.5	16.0	31.6	8.5
ZSON [ 38 ]	$\checkmark$	$\times$	$\times$	25.5	12.6	15.3	4.8

TABLE III: Comparison with SOTA methods on HM3D and MP3D datasets (1000 episodes) for object goal navigation task. The best results and second best are in bold and underlined , respectively.

-

represents that the corresponding data is not in the original report.

$o_{f}(F)$ footprint	$o_{d}(F)$ distance	$o_{s}(F)$ semantics	SR( $\%$ ) $\uparrow$	SPL( $\%$ ) $\uparrow$
$\times$	✓	✓	54.3	21.4
✓	$\times$	✓	56.7	22.8
✓	✓	$\times$	56.2	23.9
✓	✓	✓	58.2	24.5

TABLE IV: Ablation study on probability map on HM3D ObjectNav.

Object-Goal Navigation

Object-goal navigation studies how robots find and approach requested objects in unfamiliar environments using visual, spatial, and language cues.

See all 82 papers

OmniNav: Robust Long-Horizon Target Navigation in Dynamic Environments
Sep 8, 2026Yujie Tang, Meiling Wang, Jinhao Jiang +4Robot Navigation Object-Goal Navigation
School of Automation, Beijing Institute of Technology, Beijing, China.
CGFM-Nav: Cognitive Graph-Field Memory for Semantic-Guided Lifelong Multimodal Embodied Navigation
Aug 29, 2026Yuxiang Xiao, Xibei Chen, Xin Zhou +3Embodied Navigation Vision-Language Navigation
Department of Mechanical Engineering, National University of Singapore, Singapore
SSMG-Nav: Enhancing Lifelong Object Navigation with Semantic Skeleton Memory Graph
Mar 2, 2026Haochen Niu, Lantao Zhang, Xingwu Ji +3Robot Navigation Object-Goal Navigation
Brain-Inspired Application Technology Center (BATC), Shanghai Jiao Tong University, Shanghai 200240, China

Sep 8, 2026J/K move · Enter open · S save

OmniNav: Robust Long-Horizon Target Navigation in Dynamic Environments

Yujie Tang, Meiling Wang, Jinhao Jiang, Sibo Zuo+3

School of Automation, Beijing Institute of Technology, Beijing, China.

Robot Navigation Object-Goal Navigation

Open paper AskPDF arXiv

Long-horizon target navigation requires a robot to sustain task execution across evolving observations, decisions, and physical interactions. This requires three coupled capabilities: maintaining valid scene memory, revising target beliefs under partial observability, and selecting interaction-feasible navigation endpoints. However, the state underlying each capability is only conditionally valid: scene representations become stale when objects move or disappear, unsuccessful searches alter beliefs over target locations, and geometrically convenient endpoints may still be infeasible for manipulation. To address these challenges, we present OmniNav, which formulates long-horizon navigation as continual inference over a factorized task state posterior coupling scene validity, target belief, and interaction feasibility. For representation, OmniNav incrementally constructs an updatable 3D object scene memory, preventing stale scene evidence from propagating to subsequent decisions. For exploration, it introduces an evidence-aware Bayesian belief-revision mechanism that derives dependency-aware region priors from semantic context, incorporates unsuccessful searches as negative evidence, and updates them for posterior-guided frontier selection. For interaction, OmniNav incorporates manipulation reachability and collision constraints into navigation-endpoint selection and propagates execution feedback through hierarchical closed-loop recovery. Extensive experiments demonstrate that OmniNav achieves the highest success rates among the compared methods on semantic ObjectNav and fine-grained instance navigation benchmarks, remains robust to target relocation, and improves real-world pick-and-place success from 53.3% to 71.7% over an adapted open-loop baseline. The project page of OmniNav is available at https://omni-nav.github.io/.

Abstract

Object-Goal Navigation

Explore similar work

OmniNav: Robust Long-Horizon Target Navigation in Dynamic Environments

CGFM-Nav: Cognitive Graph-Field Memory for Semantic-Guided Lifelong Multimodal Embodied Navigation

SSMG-Nav: Enhancing Lifelong Object Navigation with Semantic Skeleton Memory Graph