Paper deep dive
VTM-Nav: Hierarchical Visual-Topological Memory for Cross-Episode Object-Goal Navigation
Xiaoran Xu, Yupeng Wu, Tianyu Xue, Yifan Xu, Xuanran Dong, Xiaoshan Yang, Changsheng Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 7/17/2026, 6:51:52 AM
Summary
The paper introduces VTM-Nav, a training-free navigation framework that leverages a persistent hierarchical Visual-Topological Memory (VTM) to enable cross-episode object-goal navigation. Unlike standard episodic protocols, VTM-Nav retains self-acquired scene experience across repeated visits to the same environment, using coarse-to-fine matching to retrieve room and object-level knowledge. It incorporates a conservative execution guard to prevent oscillations and premature stopping. Evaluated on HM3D v0.1, HM3D v0.2, and MP3D benchmarks, VTM-Nav outperforms the WMNav baseline, demonstrating robust experience reuse without model parameter updates.
Entities (11)
Relation Signals (10)
VTM-Nav → evaluatedon → HM3D v0.1
confidence 99% · evaluate VTM-Nav on three benchmarks, HM3D v0.1, HM3D v0.2, and MP3D
VTM-Nav → evaluatedon → MP3D
confidence 99% · evaluate VTM-Nav on three benchmarks, HM3D v0.1, HM3D v0.2, and MP3D
VTM-Nav → evaluatedon → HM3D v0.2
confidence 99% · evaluate VTM-Nav on three benchmarks, HM3D v0.1, HM3D v0.2, and MP3D
VTM-Nav → addresses → Cross-Episode Object-Goal Navigation
confidence 97% · We introduce Cross-Episode Object-Goal Navigation... To support experience reuse, we present VTM-Nav
VTM-Nav → comparedwith → WMNav
confidence 96% · compare it with a strengthened WMNav baseline augmented with cross-episode textual memory
VTM-Nav → implements → Visual-Topological Memory (VTM)
confidence 95% · we present VTM-Nav, a training-free VLM navigation framework with a persistent hierarchical Visual-Topological Memory (VTM).
VTM-Nav → utilizes → Vision-Language Model (VLM)
confidence 94% · training-free VLM navigation framework
VTM-Nav → incorporates → Conservative Execution Guard
confidence 93% · A conservative execution guard further mitigates oscillations, blocked motions, and premature stopping.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Object-goal navigation requires an embodied agent to locate and reach an instance of a specified object category in an indoor environment. Recent training-free approaches leverage vision-language models (VLMs) for open-vocabulary semantic reasoning, but are typically evaluated under an episodic protocol that resets all scene-specific state after each episode. We introduce Cross-Episode Object-Goal Navigation, in which an agent repeatedly operates in the same scene, retains only self-acquired experience, and keeps its model parameters fixed. To support experience reuse, we present \method, a training-free VLM navigation framework with a persistent hierarchical Visual-Topological Memory (VTM). The VTM organizes scene knowledge at room and object levels and retrieves relevant experience through coarse-to-fine matching, providing memory as soft guidance only when it agrees with current observations. A conservative execution guard further mitigates oscillations, blocked motions, and premature stopping. Under a controlled same-scene protocol, we evaluate \method{} on three benchmarks, HM3D v0.1, HM3D v0.2, and MP3D, and compare it with a strengthened WMNav baseline augmented with cross-episode textual memory, while keeping the VLM backbone and action pipeline identical. \method{} achieves the best performance across all three benchmarks, demonstrating the effectiveness and robustness of structured visual-topological experience reuse across datasets.
Tags
Links
- Source: https://arxiv.org/abs/2607.14514v1
- Canonical: https://arxiv.org/abs/2607.14514v1
Trouble viewing inline? Open PDF directly →
Full Text
58,029 characters extracted from source content.
Expand or collapse full text
VTM-Nav: Hierarchical Visual-Topological Memory for Cross-Episode Object-Goal Navigation Xiaoran Xu∗ , Yupeng Wu∗ , Tianyu Xue , Yifan Xu , Xuanran Dong , Xiaoshan Yang† , and Changsheng Xu X. Xu, Y. Wu, X. Yang, Y. Xu, and C. Xu are with the MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China, and also with the School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, Beijing, China (e-mail: xuxiaoran22@mails.ucas.ac.cn; wuyupeng23@mails.ucas.cn; yifan.xu, xiaoshan.yang, changsheng.xu@nlpr.ia.ac.cn).T. Xue is an undergraduate student at Tsinghua University, Beijing, China(e-mail:xuety23@mails.tsinghua.edu.cn). X. Dong is an undergraduate student at the University of Chinese Academy of Sciences, Beijing, China(e-mail:dongxuanran24@mails.ucas.ac.cn).∗Equal contribution. †Corresponding author. Abstract Object-goal navigation requires an embodied agent to locate and reach an instance of a specified object category in an indoor environment. Recent training-free approaches leverage vision-language models (VLMs) for open-vocabulary semantic reasoning, but are typically evaluated under an episodic protocol that resets all scene-specific state after each episode. We introduce Cross-Episode Object-Goal Navigation, in which an agent repeatedly operates in the same scene, retains only self-acquired experience, and keeps its model parameters fixed. To support experience reuse, we present VTM-Nav, a training-free VLM navigation framework with a persistent hierarchical Visual-Topological Memory (VTM). The VTM organizes scene knowledge at room and object levels and retrieves relevant experience through coarse-to-fine matching, providing memory as soft guidance only when it agrees with current observations. A conservative execution guard further mitigates oscillations, blocked motions, and premature stopping. Under a controlled same-scene protocol, we evaluate VTM-Nav on three benchmarks, HM3D v0.1, HM3D v0.2, and MP3D, and compare it with a strengthened WMNav baseline augmented with cross-episode textual memory, while keeping the VLM backbone and action pipeline identical. VTM-Nav achieves the best performance across all three benchmarks, demonstrating the effectiveness and robustness of structured visual-topological experience reuse across datasets. 1 Introduction Object-goal navigation (ObjectNav) requires an embodied agent to navigate in an indoor scene, find an instance of a target object category, and stop near it [2, 4]. Recent training-free agents built on vision-language models (VLMs) have made ObjectNav substantially more flexible by exploiting open-vocabulary semantics and object-room priors without task-specific training [11, 34, 21, 24]. Despite this progress, the dominant evaluation protocol remains episode-isolated: each run starts from a random pose, requests one target category, and discards all scene knowledge once the episode ends [2, 4]. This assumption is convenient for benchmarking, but it mismatches deployment. A service robot operating in a home, office, or hotel will revisit the same scene across many tasks, and should progressively exploit previously discovered room connectivity, informative object viewpoints, remote visual hints, and routes that already led to targets. Under the standard protocol, however, an agent re-enters a familiar scene as if it had never seen it before. Figure 1: Cross-episode ObjectNav in a repeatedly visited scene. Under the standard episode-isolated protocol, all acquired scene knowledge is discarded after each episode, forcing the agent to repeatedly rediscover room connectivity and target-relevant visual evidence. In contrast, VTM-Nav maintains a scene-scoped hierarchical Visual-Topological Memory that incrementally accumulates room-level topology and room-conditioned object experience, and retrieves this knowledge to guide subsequent episodes in the same scene. As illustrated in Fig. 1, we study a cross-episode ObjectNav setting that makes this deployment scenario explicit. A scene hosts a sequence of navigation episodes, each with its own start pose and target category. The target vocabulary and model parameters remain fixed, while a scene-specific experience state evolves across episodes using only information acquired through the agent’s own observations and actions. This setting enables us to examine whether accumulated scene experience improves navigation success in the same environment. This setting is related to, but distinct from, existing memory-augmented navigation. Many recent methods build memory for localization, exploration, or planning within a single episode, such as working memory, global-to-ego memory, or online topological representations [18, 37, 13, 1, 21]. A separate line has begun to study repeated or lifelong navigation with persistent memory [14, 22]. Our goal is narrower and more controlled: we isolate the value of cross-episode experience reuse in same-scene ObjectNav under fixed weights, and without ground-truth maps or oracle labels in the decision loop. We also distinguish our setting from Continual-ObjectNav, where goal categories evolve across stages and the central issue is continual adaptation under potential forgetting [36]. A central design question is what experience should persist and how it should be organized. Dense metric maps retain detailed geometry but may be unnecessarily complex and brittle under noisy open-vocabulary perception, whereas purely textual memory does not explicitly capture scene connectivity or reusable viewpoints. We therefore represent accumulated experience using a hierarchical Visual-Topological Memory that couples coarse room-level topology with fine-grained, room-conditioned object experience. Based on this idea, we propose VTM-Nav, a training-free VLM framework for cross-episode ObjectNav. For each scene, VTM-Nav maintains a persistent VTM populated only from the agent’s egocentric observations and navigation outcomes. During navigation, memory retrieval proceeds hierarchically: room-level topology is matched first to propose plausible room candidates, and object-level topology is then matched within those rooms to ground target-specific evidence and reusable approach records. Retrieved experience is injected as soft context and may bias candidate actions only when supported by the current visual observation, so that memory reuse complements rather than overrides the base VLM policy. To prevent persistent memory from lengthening trajectories through stale or weakly grounded cues, we further introduce a conservative execution guard that mitigates no-progress behavior, oscillations, blocked approaches, and premature stopping, while preserving target-visible and high-confidence actions. We evaluate VTM-Nav under a controlled same-scene protocol on HM3D v0.1, HM3D v0.2, and MP3D. We compare against both WMNav and a strengthened WMNav baseline augmented with scene-scoped cross-episode textual memory, while keeping the VLM backbone, candidate-action pipeline, action space, and low-level controller identical. This controlled design isolates the contribution of structured visual-topological experience reuse from differences in perception or action execution. The contributions of this paper are summarized as follows: • We formulate Cross-Episode Object-Goal Navigation, in which an agent repeatedly operates in the same scene with fixed model parameters and accumulates only experience acquired through its own observations and actions. • We propose VTM-Nav, a training-free VLM navigation framework with a persistent hierarchical Visual-Topological Memory that supports coarse-to-fine retrieval of room-level topology and room-conditioned object experience. • We develop a visually grounded memory-reuse mechanism and a conservative execution guard that improve navigation reliability by reducing revisits, oscillations, blocked approaches, and premature stopping. • Across all three benchmarks, VTM-Nav outperforms both WMNav and its strengthened variant augmented with cross-episode textual memory, achieving the best performance among the evaluated methods. 2 Related Work 2.1 Training-Free Object-Goal Navigation with VLMs Object-goal navigation requires an embodied agent to find an instance of a requested object category and stop near it in an indoor scene [2, 4]. Earlier ObjectNav systems commonly relied on learned policies, semantic exploration, task-specific supervision, or offline visual representation learning [26, 32, 8]. Recent zero-shot or training-free systems instead use open-vocabulary perception, foundation models, and VLM reasoning to guide search without retraining a navigation policy [9, 34, 11, 21, 24, 6, 12]. These agents can exploit object-room priors, commonsense spatial relations, and language-conditioned visual evidence, making them a natural backbone for deployment in semantically diverse scenes. The limitation relevant to this paper is not that these systems lack semantic priors, but that they are usually evaluated as isolated episodes. After one run terminates, room connectivity, object viewpoints, and successful approach traces are discarded, even though a deployed robot may revisit the same home, office, or hotel many times. Our work therefore keeps the VLM backbone and action pipeline fixed, and asks a narrower question: whether self-acquired scene experience can improve success rate across repeated same-scene ObjectNav episodes at comparable path efficiency. 2.2 Memory and Topological Representations for Navigation Memory has long been used to address partial observability in embodied navigation. Metric and semantic maps support exploration and planning but depend on reliable localization, occupancy estimation, and object labeling [8]. Topological memories abstract an environment into places and connectivity, which is often sufficient for high-level navigation decisions [27, 1, 40]. Recent systems further introduce scene graphs, episodic memory, working memory, self-refining graph memory, and global-to-egocentric memory to retain visited states, object evidence, spatial relations, or subgoal candidates during navigation [33, 19, 39, 18, 13, 37, 21]. The distinction in VTM-Nav is the scope and organization of memory. Most memory-augmented navigation methods optimize within-episode localization, exploration, or planning, and the memory state is reset before the next benchmark episode. In contrast, our Visual-Topological Memory (VTM) is scene-scoped and persistent across episodes, storing coarse room topology together with fine-grained in-room evidence, remote-visible cues, viewpoints, and success-backed records. The novelty here is therefore not a new memory primitive but the persistence and cross-episode reuse of an otherwise standard hierarchical topological memory. 2.3 Cross-Episode, Lifelong, and Continual Navigation Recent benchmarks and systems have begun to study navigation beyond a single short episode. GOAT-Bench evaluates agents over long sequences of navigation goals, and lifelong ObjectNav work investigates how persistent semantic memory can reduce redundant exploration across repeated requests [14, 22]. These efforts share our deployment motivation: embodied agents should benefit from previous interaction instead of solving every request from scratch. Our setting is deliberately more controlled. The target vocabulary, VLM, perception stack, and action pipeline remain fixed; the only evolving state is the scene-specific experience collected by the agent itself. This separates our problem from Continual-ObjectNav, where new goal categories arrive over stages and the core challenge is learning without forgetting [36, 23, 17], and from test-time adaptation, where model parameters or representations are updated online [10, 15]. We instead study gradient-free, cross-episode experience reuse, with success-rate improvement at comparable path efficiency as the central empirical signature. 3 Method Figure 2: Overview of VTM-Nav. The upper branch illustrates the step-level online decision process. At each step, the current RGB-D observation, target category, pose, and previous action are converted into structured semantic evidence, which supports room localization through semantic, topological, and spatial consistency. The localized observation is then used to update a scene-scoped hierarchical Visual-Topological Memory containing a room-level topology, room-conditioned object nodes, and episode history. Target-conditioned retrieval produces a scene prior that summarizes the plausible target room, retrieval confidence, remembered room path, and approach hint, which are used to guide action selection. The lower branch illustrates episode-level memory evolution. After termination, room visits, transitions, object observations, and episode outcomes are written back to memory, while successful episodes additionally consolidate target-room associations, terminal viewpoints, and successful approach cues for subsequent episodes in the same scene. 3.1 Problem Formulation We study a cross-episode, experience-evolving extension of standard ObjectNav. The task definition, observation and action spaces, success criterion, target vocabulary, and model parameters are fixed. The only additional state is scene-specific experience collected by the agent itself. This setting asks whether later episodes in the same scene can benefit from earlier exploration without updating the VLM policy. In a standard ObjectNav episode, an agent starts at pose P0P_0 in an indoor environment ℰE and receives a target category g∈g . At step t, it observes OtO_t, estimates its pose PtP_t, and obtains a set of navigable candidates t=ct,kk=1Kt.C_t=\c_t,k\_k=1^K_t. (1) Each candidate ct,kc_t,k has a polar displacement (rt,k,θt,k)(r_t,k, _t,k) and a short textual description dt,kd_t,k from the perception module. The agent either selects a candidate for traversal or issues stop; an episode succeeds when it stops within dsuccd_succ of a valid target instance. For each scene s, we organize its episodes as Es=(es,1,es,2,…,es,Js),E_s=(e_s,1,e_s,2,…,e_s,J_s), (2) where episode es,je_s,j has its own start pose Ps,j,0P_s,j,0 and target category gs,jg_s,j. The agent maintains a scene memory ℳsM_s, initialized as ∅ before the first episode and retained only within the same scene. After episode es,je_s,j, the memory update is ℳs(j)=Φwrite(ℳs(j−1),τs,j),ℳs(0)=∅,M_s^(j)= _write (M_s^(j-1), _s,j ), _s^(0)= , (3) where trajectory τs,j _s,j contains the observations, poses, actions, and outcome from that episode. The memory is built only from information available during navigation: egocentric observations, estimated poses, semantic predictions, executed actions, and episode outcomes. It does not use ground-truth object locations, oracle semantic maps, oracle room labels, or shortest paths for action selection. Memory is scene-scoped, so ℳsM_s cannot be queried in another scene s′≠s ≠ s. 3.2 Overview of VTM-Nav VTM-Nav is a training-free VLM navigation framework that augments a WMNav-style backbone with a scene-scoped experience layer. The base backbone constructs navigable candidates from depth observations and uses a VLM to select an action according to the current visual evidence [21, 11]. We do not retrain or replace its VLM, candidate-generation mechanism, or low-level controller. Instead, VTM-Nav organizes experience reuse into five operations shown in Fig. 2: semantic evidence extraction, scene-topological memory maintenance, target-conditioned memory retrieval, soft memory-guided VLM action execution, and conservative guard before action execution. At time step t, VTM-Nav follows the pipeline in Fig. 2. The semantic perception block converts the current observation and target context into structured room- and object-level evidence. The memory block then localizes the agent in the stored room topology, updates room and object records, and retrieves target-conditioned scene priors. These priors are supplied to the base VLM as soft memory guidance. They can bias candidate selection only when they are reliable and consistent with current visual evidence. After memory-guided action selection, the recovery guard checks clear execution risks, including premature stopping, blocked target approaches, non-navigable candidates, no-progress motion, revisits, and short-horizon oscillation. The final episode outcome is then written back to the same scene memory, closing the cross-episode experience loop. At a high level, each scene memory contains a room topology, room-owned object memories, and episode-level search history. The next section defines this representation; here we summarize how it is used online. Let ℳs,j,tM_s,j,t denote the scene memory before step t of episode es,je_s,j. The step-level flow is zt→λt→ℳs,j,t+→ρt→a^t→a~t→at.z_t→ _t _s,j,t^+→ _t→ a_t→ a_t→ a_t. (4) Here, ztz_t is the structured semantic evidence extracted from the current observation; λt _t is the room-localization result; ℳs,j,t+M_s,j,t^+ is the memory after writing the current room and object evidence; ρt _t is the target-conditioned scene prior retrieved from the updated memory; a^t a_t is the action proposed by the unchanged VLM policy; a~t a_t is the action after conservative memory-based candidate bias; and ata_t is the executed action after the guard check. The short-term history hth_t is used only by the guard to detect local execution failures such as no-progress motion or oscillation. At episode termination, the final trajectory and outcome are written back to the scene memory. The resulting persistent state initializes the next episode in the same scene, allowing later episodes to retrieve room topology, object evidence, and approach information acquired earlier while keeping the VLM and navigation-controller parameters fixed. 3.3 Hierarchical Visual-Topological Memory 3.3.1 Semantic Evidence The Visual-Topological Memory (VTM) stores scene experience extracted from the agent’s own observations. At each step, semantic perception produces an evidence tuple zt=(r^t,^t,ut,Pt,Rt,at−1,g),z_t=( r_t, Q_t,u_t,P_t,R_t,a_t-1,g), (5) where r^t r_t is the perceived room category, ^t Q_t is the normalized set of visible objects, utu_t is a compact visual summary, PtP_t and RtR_t denote the estimated position and orientation, at−1a_t-1 is the previous action, and g is the current target category. Before writing to memory, we normalize the semantic evidence by filtering room names from object entries, removing negated object mentions, and treating ObjectNav targets as object categories rather than room categories. Objects visible through doorways or openings are stored as remote-visible evidence instead of confirmed in-room evidence. 3.3.2 Hierarchical Memory Structure The VTM represents scene experience using a two-level room-to-object hierarchy. At the coarse level, it maintains a room-level topological graph s=(sroom,ℰsroom),T_s=(V^room_s,E^room_s), (6) where sroomV^room_s is the set of room nodes and ℰsroomE^room_s is the set of observed transitions between them. Each room node stores its canonical room label, semantic aliases, localization confidence, visit statistics, and recent visual summaries. At the fine level, each room node v∈sroomv ^room_s is associated with a local object memory s,v=(s,vin,s,vrem),O_s,v=(O^in_s,v,O^rem_s,v), (7) where the superscripts inin and remrem denote in-room and remote-visible, respectively. Specifically, s,vinO^in_s,v contains object nodes directly observed inside room v, whereas s,vremO^rem_s,v contains object cues observed from room v through doorways or other openings. The object-level memories associated with all room nodes are collected as s=s,v∣v∈sroom.O_s=\O_s,v v ^room_s\. (8) The complete scene memory is then represented as ℳs=(s,s,ℋs),M_s=(T_s,O_s,H_s), (9) where ℋsH_s stores episode-level search outcomes and navigation history. Each object node o∈s,vo _s,v records its category, confidence, observation count, representative viewpoints, visual summaries, navigation hints, and success-backed approach information when available. A remote-visible object node is associated with the room from which it was observed and additionally records the corresponding opening relation and, when inferable, a hypothesized destination room. It is therefore not treated as confirmed evidence that the object lies inside the observer room. Unlike a flat scene-level object collection, sO_s preserves the room ownership of every object node. This nested organization supports coarse-to-fine experience reuse: the agent first retrieves a plausible target-bearing room from sT_s and then queries the object nodes in the corresponding local memory s,vO_s,v. 3.3.3 Room Localization Given the current evidence ztz_t, VTM first localizes the agent to a room node. For each candidate room node v∈sroomv ^room_s, we compute a bounded localization score Sloc(v,zt)=Ssem(v,zt)+Stopo(v,zt)+Ssp(v,zt),S_loc(v,z_t)=S_sem(v,z_t)+S_topo(v,z_t)+S_sp(v,z_t), (10) where SlocS_loc is the room-localization score. The three components compare the current evidence ztz_t with the stored room node v: SsemS_sem measures semantic consistency, StopoS_topo measures topological consistency, and SspS_sp measures spatial consistency. The semantic consistency score uses the room name, visible objects, and visual summary contained in ztz_t: Ssem(v,zt)=13(Sname(v,zt)+Sovl(v,zt)+Sdesc(v,zt)).S_sem(v,z_t)= 13 (S_name(v,z_t)+S_ovl(v,z_t)+S_desc(v,z_t) ). (11) To make these terms explicit, let vA_v be the canonical room name and aliases of node v. Let vC_v be its stored object-category set, and let vU_v be its recent visual summaries. With r^t r_t, ^t Q_t, and utu_t taken from ztz_t, we use Sname(v,zt) S_name(v,z_t) =[r^t∈v], =I[ r_t _v], (12) Sovl(v,zt) S_ovl(v,z_t) =J(^t,v), =J( Q_t,C_v), Sdesc(v,zt) S_desc(v,z_t) =maxui∈vJ(kw(ut),kw(ui)). = _u_i _vJ(kw(u_t),kw(u_i)). Here, [⋅]I[·] is 1 when the condition holds and 0 otherwise. J(A,B)=|A∩B|/|A∪B|J(A,B)=|A∩ B|/|A∪ B| is the Jaccard overlap, and is set to 0 when A∪BA∪ B is empty. kw(⋅)kw(·) extracts normalized keywords from a visual summary, and SdescS_desc is set to 0 when vU_v is empty. When forming vC_v, confirmed in-room objects are used before remote-visible objects; if no confirmed object category is stored for v, remote-visible categories provide a fallback cue. The topological score uses the previous localized room vt−1⋆v_t-1 : Stopo(v,zt)=[v=vt−1⋆∨(vt−1⋆,v)∈ℰsroom],S_topo(v,z_t)=I [v=v_t-1 \ \ (v_t-1 ,v) ^room_s ], (13) where [⋅]I[·] is 1 when the condition holds and 0 otherwise. The spatial score compares the current pose PtP_t with recent poses assigned to v: Ssp(v,zt)=maxPi∈v11+∥Pt−Pi∥2,S_sp(v,z_t)= _P_i _v 11+ P_t-P_i _2, (14) where vP_v is the set of recent pose estimates stored for room v; if vP_v is empty, the spatial score is set to 0. The best-matching room is selected by vt⋆=argmaxv∈sroomSloc(v,zt).v_t = _v ^room_sS_loc(v,z_t). (15) Here, vt⋆v_t is assigned to the current observation when its localization score is sufficiently high. Otherwise, the observation is kept as tentative evidence. A new room node is promoted only after repeated tentative support or a high-confidence room observation, which prevents transient VLM room descriptions from immediately polluting the persistent topology. This localization output corresponds to λt _t in the step-level flow. 3.3.4 Object-Node Association After room localization, each object observation is associated with the local object memory of the matched room vt⋆v_t . We first justify whether an observed object q∈^tq∈ Q_t is directly visible inside the current room or visible through a doorway or another opening. This yields an evidence type κt(q)∈in,rem, _t(q)∈\in,rem\, (16) which determines whether the observation is written to s,vt⋆inO^in_s,v_t or s,vt⋆remO^rem_s,v_t . The matched object node is ot⋆(q):=argmaxo∈s,vt⋆κt(q)Sobj(o,q),o_t (q):= _o _t(q)_s,v_t S_obj(o,q), (17) where Sobj(o,q)S_obj(o,q) measures category agreement and compatibility between the current and stored visual descriptions. If the best matching score exceeds a fixed threshold τobj _obj, the observation updates ot⋆(q)o_t (q). Otherwise, a new object node is created under vt⋆v_t . Object nodes are therefore created and updated locally within the room from which their evidence was acquired. For a remote-visible object, the node is attached to the observer room vt⋆v_t and additionally records the corresponding doorway or opening cue and, when inferable, a hypothesized destination room. Such a node is not treated as confirmed evidence that the object lies inside the observer room. 3.3.5 Memory Update After room localization, the VTM updates the corresponding room, edge, and object records. If the localization result matches an existing room, the room node updates its visit count, localization confidence, and recent visual summaries using the current evidence. If tentative room evidence is promoted to a new room node vnewv_new, the room-node set is updated as sroom←sroom∪vnew.V^room_s ^room_s∪\v_new\. (18) When a confirmed transition from room viv_i to room vjv_j is observed, the corresponding edge enewe_new is inserted if absent, shown as ℰsroom←ℰsroom∪enew.E^room_s ^room_s∪\e_new\. (19) And its statistics are updated as γij←min(1,γij+Δe),nij←nij+1. _ij← (1, _ij+ _e), n_ij← n_ij+1. (20) For an in-room object entry o, each direct observation increases its confidence: γo←min(1,γo+Δo0). _o← (1, _o+ _o^0). (21) At the end of a successful episode, the object entry associated with the final target approach is marked as success-backed and receives an additional update: γo←min(1,γo+Δo+),Δo+>Δo0>0. _o← (1, _o+ _o^+), _o^+> _o^0>0. (22) Its viewpoint, visual summary, and approach direction are stored with the object record. The updated records define ℳs,j,t+M_s,j,t^+ for the current step. 3.4 Memory-Guided Action Selection 3.4.1 Cross-Episode Retrieval Given a target category g, the memory module retrieves both confirmed in-room evidence and remote-visible evidence from ℳsM_s. For a room v, in-room evidence is scored by Rin(v,g)=γv,g+η1nv,g+η2nv,gsucc,R_in(v,g)= _v,g+ _1n_v,g+ _2n^succ_v,g, (23) where γv,g _v,g is the stored confidence of target g in room v, nv,gn_v,g is how many times it has been observed there, and nv,gsuccn^succ_v,g is the number of episodes in which the agent actually reached g from that room. The two coefficients η1,η2 _1, _2 are small fixed weights, so the score is dominated by confidence and only refined by repeated and success-backed evidence. Remote-visible evidence is scored with the same fields, but its confidence is capped below confirmed in-room evidence. This gives in-room records priority while still letting a target glimpsed through a doorway guide the agent toward the correct neighboring room. The path is computed over the stored topology: If the current room vtv_t and the retrieved room vgv_g are connected, we use breadth-first search (BFS) on the stored room graph to obtain the shortest observed room-hop path: πt=BFS(s,vt,vg). _t=BFS (T_s,v_t,v_g ). (24) If no connection has been observed, the path field is left empty. The resulting prior ρt _t contains the evidence type, relevant room, confidence, viewpoint or approach hint, and the optional room path. 3.4.2 Candidate Bias The prior ρt _t is first inserted into the VLM prompt as soft context, so the navigation policy can already account for remembered evidence without being forced to follow it. On top of this, we add a lightweight, deterministic check that re-ranks the proposed navigable candidates using the same retrieved memory. Let k^t k_t be the candidate index selected by the VLM action a^t a_t. Using the candidate descriptions dt,kd_t,k defined above, we read off from ρt _t a set of positive cues, including the remembered target room name, the target object name, and the stored approach or viewpoint hint, together with a fixed set of negative cues that signal non-navigable structure. Each candidate then receives an agreement score Bt,k=clip(ψmem(dt,k,ρt)+ψnav(dt,k)−ψblock(dt,k), 0,1),B_t,k=clip ( _mem(d_t,k, _t)+ _nav(d_t,k)- _block(d_t,k),\;0,1 ), (25) where every term is computed by simple lexical matching rather than a learned model, so the rule is transparent and reproducible. The memory term ψmem _mem rewards a candidate whose description overlaps the positive cues, with a stronger reward when it explicitly names the remembered target room or target object. The navigational term ψnav _nav rewards traversal cues such as an open doorway, corridor, or passage that tend to lead toward other rooms. The blocking term ψblock _block penalizes non-navigable cues such as a closed door, wall, dead end, or staircase. Let kt†=argmaxkBt,k _t= _kB_t,k be the highest-scoring candidate. The biased action selects k~t=kt† k_t=k _t only when Bt,kt†B_t,k _t clears a single confidence threshold; otherwise k~t=k^t k_t= k_t, i.e. the action originally chosen by the VLM is kept. We write the corresponding action as a~t a_t. This makes memory reuse conservative: it can promote a memory-consistent direction among the existing candidates, but it never creates a waypoint outside the current action space. 3.4.3 Conservative Execution Guard Memory bias uses long-term scene experience, but it does not detect short-term failures such as no-progress motion, repeated visits, or left–right/forward–back oscillation. We therefore add a guard that stays inactive by default and only reviews the selected candidate when such signals appear. When triggered, the guard adds a motion-efficiency term to the current candidate preference: Sguard(k)=Sbase(k)+αSeff(k).S_guard(k)=S_base(k)+α S_eff(k). (26) Here, SbaseS_base contains the VLM, memory-bias, and traversability preferences, while SeffS_eff rewards forward progress and penalizes revisits, backtracking, oscillation, and dead ends. The guard replaces the memory-guided candidate k~t k_t with kt⋆k_t only if Sguard(kt⋆)≥Sguard(k~t)+δguard.S_guard(k_t )≥ S_guard( k_t)+ _guard. (27) It otherwise leaves the original choice unchanged. The guard also never overrides stop when the target is visible, so it remains a last-resort correction rather than a second planner. The resulting choice is the executed action ata_t. TABLE I: Object-goal navigation results on HM3D. TF and ZS denote training-free and zero-shot settings, respectively. Steps denotes the maximum number of interaction steps allowed per episode. Rows marked with ∗ are our reproductions of WMNav, because the original WMNav paper used Gemini-1.5-Pro, which is no longer available. “–” denotes unreported entries. Method Foundation model TF ZS Steps HM3D v0.1 HM3D v0.2 SR ↑ SPL ↑ SR ↑ SPL ↑ Task-trained and non-zero-shot navigation methods Habitat-Web [26] – × × 500 41.5 16.0 – – OVRL-V2 [31] – × × 500 64.7 28.1 – – Zero-shot but task-trained navigation methods ZSON [20] CLIP × ✓ 500 25.5 12.6 – – PSL [29] CLIP × ✓ 500 42.4 19.2 – – PixNav [5] Foundation models × ✓ 500 37.9 20.5 – – SGM [38] – × ✓ 500 60.2 30.8 – – VLFM [34] BLIP-2 × ✓ 500 52.5 30.4 62.6 31.0 Training-free and zero-shot navigation methods ESC [41] GLIP-L + DeBERTa-v3 ✓ ✓ 500 39.2 22.3 – – L3MVN [35] GPT-2 Large ✓ ✓ 500 50.4 23.1 36.3 15.7 VoroNav [30] GPT-3.5 ✓ ✓ 500 42.0 26.0 – – OpenFMNav [16] GPT-4/4V + G-SAM ✓ ✓ 500 54.9 24.4 – – WMNav∗ [21] Qwen3-vl-Plus ✓ ✓ 40 55.0 31.7 70.0 30.0 Cross-episode experience-evolving navigation methods WMNav + WMNav + Textual Memory∗ Qwen3-vl-Plus ✓ ✓ 40 56.5 31.1 66.5 31.2 VTM-Nav Qwen3-vl-Plus ✓ ✓ 40 59.6 31.8 72.0 31.5 VTM-Nav Qwen3-vl-Plus ✓ ✓ 500 65.3 32.1 – – TABLE I: Object-goal navigation results on MP3D. TF and ZS denote training-free and zero-shot settings, respectively. Rows marked with ∗ are our reproductions of WMNav, because the original WMNav paper used Gemini-1.5-Pro, which is no longer available. Method Foundation model TF ZS Steps SR ↑ SPL ↑ Task-trained and non-zero-shot navigation methods Habitat-Web [26] – × × 500 31.6 8.5 OVRL [32] – × × 500 28.6 7.4 Zero-shot but task-trained navigation methods ZSON [20] CLIP × ✓ 500 15.3 4.8 PSL [29] CLIP × ✓ 500 18.9 6.4 SGM [38] – × ✓ 500 37.7 14.7 VLFM [34] BLIP-2 × ✓ 500 36.4 17.5 Training-free and zero-shot navigation methods CoW [9] CLIP ✓ ✓ 500 9.2 4.9 ESC [41] GLIP-L + DeBERTa-v3 ✓ ✓ 500 28.7 14.2 WMNav∗ [21] Qwen3-vl-Plus ✓ ✓ 40 43.5 15.6 Cross-episode experience-evolving navigation methods VTM-Nav Qwen3-vl-Plus ✓ ✓ 40 44.3 16.2 4 Experiments 4.1 Datasets and Evaluation Metrics Benchmarks. We evaluate VTM-Nav on Habitat ObjectNav benchmarks using HM3D v0.1, HM3D v0.2, and MP3D [28, 7, 25, 4]. HM3D v0.1 is the Habitat 2022 ObjectNav benchmark, with 2000 validation episodes over 20 validation scenes and 6 goal categories. HM3D v0.2 uses a cleaned version of HM3D with improved geometry and semantic annotations, and contains 1000 validation episodes. MP3D provides 11 validation scenes, 2195 validation episodes, and 21 object-goal categories. We use validation splits and scene-level sharding in all evaluations. Metrics. We report Success Rate (SR) and Success weighted by Path Length (SPL) [2]. For N evaluation episodes, SR=1N∑i=1NSi,SR= 1N _i=1^NS_i, (28) where Si∈0,1S_i∈\0,1\ indicates whether episode i is successful. SPL is defined as SPL=1N∑i=1NSiℓi⋆max(ℓi,ℓi⋆),SPL= 1N _i=1^NS_i _i ( _i, _i ), (29) where ℓi _i is the executed path length and ℓi⋆ _i is the shortest geodesic path to a valid goal location. We also log memory use, loop behavior, path length, and guard interventions to diagnose where the gain comes from. Cross-episode protocol. To isolate persistent scene experience, all episodes from the same scene are assigned to the same worker and executed consecutively. The worker maintains a scene-specific Visual-Topological Memory (VTM), which is initialized before the first episode of a scene, updated during navigation, and retained for later episodes in that scene. Memory is never shared across scenes or workers. For the episode-isolated counterpart, we keep the same worker assignment, scene order, start poses, target categories, action space, and step budget, but reset the scene memory before every episode. 4.2 Implementation Details We adapt WMNav∗ and a strengthened baseline, denoted as WMNav + Textual Memory∗. WMNav∗ is our reimplementation of WMNav under the controlled evaluation pipeline. WMNav + Textual Memory∗ further equips WMNav with a scene-scoped cross-episode textual memory: after each episode, the VLM summarizes navigation outcomes, including successful strategies and failure cases, and retains these summaries for use in subsequent episodes within the same scene. This baseline allows us to distinguish the benefit of persistent experience in general from that of the proposed hierarchical visual-topological representation. All methods mentioned above use the same Qwen3-VL-Plus backbone [3], candidate-action pipeline, Habitat ObjectNav action space, and low-level controller. Following the WMNav configuration, the agent has a cylindrical body with a radius of 0.18 m and a height of 0.88 m. It receives egocentric RGB-D observations at a resolution of 640×480640× 480 with a 79∘79 horizontal field of view. The camera is pitched 14∘14 downward to support navigability estimation, and each episode is limited to 40 interaction steps. VTM-Nav maintains one scene-scoped memory for each worker and updates it after every episode. All retrieval thresholds and execution-guard margins are fixed across datasets. 4.3 Comparison with SOTA Methods Tables I and I compare VTM-Nav with prior ObjectNav systems under several deployment settings. Task-trained and non-zero-shot methods use ObjectNav training data and a fixed target vocabulary. Zero-shot but task-trained methods can generalize to unseen categories, but still rely on task-specific navigation training. Training-free and zero-shot methods do not update navigation parameters on ObjectNav and instead use foundation-model priors at test time. On HM3D, VTM-Nav consistently outperforms the controlled baselines. On HM3D v0.1, it reaches 59.6 SR, improving over WMNav∗ by 4.6 points and over WMNav + WMNav + Textual Memory by 3.1 points. With a longer 500-step budget, VTM-Nav further reaches 65.3 SR and 32.1 SPL on HM3D v0.1, showing that the accumulated memory can still help when the interaction budget is relaxed. On HM3D v0.2, VTM-Nav reaches 72.0 SR, improving over WMNav∗ by 2.0 points and over WMNav + Textual Memory∗ by 5.5 points. SPL remains comparable across the controlled 40-step variants, showing that the success gains are achieved without substantially lengthening successful trajectories. On MP3D, VTM-Nav also improves both metrics over WMNav∗, increasing SR from 43.5 to 44.3 and SPL from 15.6 to 16.2. Across datasets, VTM-Nav improves over the WMNav backbone under the same VLM and 40-step action budget. Compared with free-form textual evolving memory, VTM provides a more reusable experience structure by tying object evidence to room ownership, observed transitions, and success-backed target records. TABLE I: Ablation study of experience modules. “Topo.” is topological structure, “Persist.” is cross-episode persistence, “Succ.” is success-backed retrieval, and “Visual” is visible target evidence. Variant Topo. Persist. Succ. Visual SR SPL WMNav backbone ✗ ✗ ✗ ✗ 55.0 31.7 w/o topological structure ✗ ✓ ✓ ✓ 56.5 31.1 w/o cross-episode persistence ✓ ✗ ✓ ✓ 59.4 30.5 w/o success-backed retrieval ✓ ✓ ✗ ✓ 58.0 30.0 w/o visible evidence ✓ ✓ ✓ ✗ 57.7 29.4 Full VTM-Nav ✓ ✓ ✓ ✓ 59.6 31.8 TABLE IV: Paired outcome and trajectory diagnostics on HM3D v0.2. Outcome metrics use all 1000 paired episodes, while trajectory metrics use the subset where both VTM-Nav and WMNav∗ succeed. Relative gains are computed against WMNav∗. Metric Scope VTM-Nav WMNav∗ Gain Success rate (%) ↑ All 72.0 70.0 +2.0% Method-only successes (#) ↑ All 112 92 +20 Backtrack count (/episode) ↓ Both succ. 1.4868 1.5263 2.6% lower Dead-end waste (steps/episode) ↓ Both succ. 3.1053 3.2105 3.3% lower Avg. progress (m/step) ↑ Both succ. 0.7959 0.7823 1.7% higher 4.4 Ablation Study Table I isolates the main memory components on HM3D v0.1 under the same 40-step protocol. The WMNav backbone removes all scene experience, while the other variants disable one part of VTM at a time. This design tests whether the gain comes from structured memory organization, cross-episode persistence, success-backed target records, or direct visible evidence. The full model improves SR from 55.0 to 59.6 over the WMNav backbone while keeping SPL nearly unchanged. Removing topological structure gives the largest SR drop among the memory variants, from 59.6 to 56.5. This indicates that reusable experience is most useful when object evidence is tied to room ownership and observed transitions, rather than stored as a loose textual record. Removing cross-episode persistence keeps SR close to the full model but reduces SPL to 30.5, which is consistent with persistent scene memory shortening successful routes after earlier exploration. Success-backed retrieval and visible target evidence mainly affect path efficiency. Without success-backed records, SPL drops to 30.0, and without visible evidence it drops further to 29.4. These results suggest that the memory does not only help choose a plausible room; it also helps decide when a retrieved target cue is reliable enough to influence the final approach. Table IV further reports paired HM3D v0.2 diagnostics. Across 1000 paired episodes, VTM-Nav reaches 72.0% SR, compared with 70.0% for WMNav∗, and converts 20 additional episodes into method-only successes. On the subset where both methods succeed, VTM-Nav also shows fewer backtracks, less dead-end waste, and higher average progress per step. The trajectory margins are modest, but they move in the same direction as the SR gain and support the claim that VTM makes the search more directed. 4.5 Cross-Episode Progression Analysis The main results compare methods over the full validation set, but they do not directly show whether persistent memory becomes more useful as the same scene is visited repeatedly. We therefore perform a within-scene progression analysis. For each scene, episodes are ordered by execution time and split into early and late halves. We then compute the late-minus-early change in SR and SPL for each scene and average the change over scenes. TABLE V: Within-scene early-to-late progression on HM3D v0.2. Episodes in each scene are split into early and late halves. Values report averages over scenes, and Δ denotes late minus early. Method SR (%) ↑ SPL (%) ↑ Early Late Δ Early Late Δ WMNav∗ 70.0 70.7 +0.7 28.8 28.2 -0.6 WMNav + Textual Memory∗ 67.7 68.0 +0.3 31.2 31.1 -0.0 VTM-Nav 70.8 73.4 +2.7 31.3 31.6 +0.3 Table V measures whether performance changes as more episodes are observed in the same scene. For each scene, we split its ordered episodes into early and late halves and average both halves across scenes. VTM-Nav increases from 70.8% to 73.4% SR, a 2.7-point gain, while keeping SPL. WMNav∗ and textual evolving memory show smaller SR gains, and WMNav∗ loses SPL in the late half. This diagnostic supports the cross-episode use of structured VTM. HM3D v0.2, plant target MP3D, chair target Figure 3: Failure-to-success comparisons on HM3D v0.2 and MP3D. In both rows, VTM-Nav reaches the target-bearing region and stops successfully, whereas WMNav continues searching and fails within the budget. Figure 4: Path-efficiency comparison under successful outcomes. Both methods reach the chair target, but VTM-Nav succeeds in 2 steps, whereas WMNav requires 40 steps. 4.6 Qualitative Visualization and Case Studies We visualize first-person observations together with top-down trajectories to show how memory changes search behavior. In each paired comparison, the left panel shows VTM-Nav and the right panel shows the WMNav backbone under the same target category and scene context. We include two failure-to-success cases and one success-to-success case, so the comparison covers both target discovery and path efficiency. Figure 3 shows the main qualitative difference behind the SR gain. In both scenes, WMNav continues to expand the explored area after several local observations fail to reveal the target. VTM-Nav instead uses the retrieved room-level prior to move toward the target-bearing region earlier. The final stop is still grounded in current visual evidence, but memory reduces the amount of unguided search needed before that evidence appears. Figure 4 isolates path efficiency. Both methods find the target, but their routes differ sharply. VTM-Nav chooses the target-bearing direction immediately and stops after 2 steps, while WMNav spends most of the budget exploring before stopping. This case explains why persistent VTM can help SPL even when the baseline eventually succeeds: the stored scene structure turns a successful but late search into a short directed trajectory. 5 Conclusion We introduced VTM-Nav, a cross-episode ObjectNav system that stores self-acquired scene experience as hierarchical Visual-Topological Memory. By linking room topology, object evidence, remote-visible cues, and success-backed target records, VTM-Nav provides a fixed VLM navigator with scene-specific priors without model training. Under the controlled 40-step protocol, it improves SR over WMNav∗ on HM3D v0.1, HM3D v0.2, and MP3D, and outperforms free-form textual evolving memory on HM3D. The results suggest that training-free VLM agents can benefit from structured scene memory, while cross-scene transfer and learned recovery remain open problems. References [1] D. An, H. Wang, W. Wang, Z. Wang, Y. Huang, K. He, and L. Wang (2025) ETPNav: evolving topological planning for vision-language navigation in continuous environments. IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (7), p. 5130–5145. External Links: Document Cited by: §1, §2.2. [2] P. Anderson, A. X. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V. Koltun, J. Kosecka, J. Malik, R. Mottaghi, M. Savva, and A. R. Zamir (2018) On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757. Cited by: §1, §1, §2.1, §4.1. [3] S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. (2025) Qwen3-VL technical report. External Links: 2511.21631, Document, Link Cited by: §4.2. [4] D. Batra, A. Gokaslan, A. Kembhavi, O. Maksymets, R. Mottaghi, M. Savva, A. Toshev, and E. Wijmans (2020) ObjectNav revisited: on evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171. Cited by: §1, §1, §2.1, §4.1. [5] W. Cai, S. Huang, G. Cheng, Y. Long, P. Gao, C. Sun, and H. Dong (2023) Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill. External Links: 2309.10309, Document, Link Cited by: TABLE I. [6] Y. Cai, X. He, M. Wang, H. Guo, W. Yau, and C. Lv (2025) CL-CoTNav: closed-loop hierarchical chain-of-thought for zero-shot object-goal navigation with vision-language models. External Links: 2504.09000, Document, Link Cited by: §2.1. [7] A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y. Zhang (2017) Matterport3D: learning from RGB-D data in indoor environments. In International Conference on 3D Vision, Cited by: §4.1. [8] D. S. Chaplot, D. Gandhi, S. Gupta, A. Gupta, and R. Salakhutdinov (2020) Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems 33, p. 4247–4258. Cited by: §2.1, §2.2. [9] S. Y. Gadre, M. Wortsman, G. Ilharco, L. Schmidt, and S. Song (2023) CoWs on pasture: baselines and benchmarks for language-driven zero-shot object navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 23171–23181. Cited by: §2.1, TABLE I. [10] J. Gao, X. Yao, and C. Xu (2024) Fast-slow test-time adaptation for online vision-and-language navigation. In Proceedings of the International Conference on Machine Learning, External Links: 2311.13209, Link Cited by: §2.3. [11] D. Goetting, H. G. Singh, and A. Loquercio (2025) End-to-end navigation with vision-language models: transforming spatial reasoning into question-answering. In Proceedings of the International Conference on Neuro-symbolic Systems, Proceedings of Machine Learning Research, Vol. 288, p. 22–35. External Links: 2411.05755, Document, Link Cited by: §1, §2.1, §3.2. [12] M. James, A. A. Habel, A. Fedoseev, and D. Tsetserokou (2026) GoalVLM: VLM-driven object goal navigation for multi-agent system. External Links: 2603.18210, Document, Link Cited by: §2.1. [13] Z. Ji, H. Lin, and Y. Gao (2025) DyNaVLM: zero-shot vision-language navigation system with dynamic viewpoints and self-refining graph memory. arXiv preprint arXiv:2506.15096. External Links: 2506.15096, Document, Link Cited by: §1, §2.2. [14] M. Khanna, R. Ramrakhya, G. Chhablani, S. Yenamandra, T. Gervet, M. Chang, Z. Kira, D. S. Chaplot, D. Batra, and R. Mottaghi (2024) GOAT-Bench: a benchmark for multi-modal lifelong navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, External Links: Link Cited by: §1, §2.3. [15] H. Ko, S. Kim, G. Oh, J. Yoon, H. Lee, S. Jang, S. Kim, and S. Kim (2025) Active test-time vision-language navigation. External Links: 2506.06630, Document, Link Cited by: §2.3. [16] Y. Kuang, H. Lin, and M. Jiang (2024) OpenFMNav: towards open-set zero-shot object navigation via vision-language foundation models. In Findings of the Association for Computational Linguistics: NAACL, External Links: 2402.10670, Document, Link Cited by: TABLE I. [17] T. Lesort, V. Lomonaco, A. Stoian, D. Maltoni, D. Filliat, and N. Diaz-Rodriguez (2020) Continual learning for robotics: definition, framework, learning strategies, opportunities and challenges. Information Fusion 58, p. 52–68. External Links: Document Cited by: §2.3. [18] H. Li, Z. Wang, X. Yang, Y. Yang, S. Mei, and Z. Zhang (2024) MemoNav: working memory model for visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 17913–17922. Cited by: §1, §2.2. [19] J. Loo, Z. Wu, and D. Hsu (2025) Open scene graphs for open-world object-goal navigation. Note: Related DOI: 10.1177/02783649251369549 External Links: 2508.04678, Document, Link Cited by: §2.2. [20] A. Majumdar, G. Aggarwal, B. Devnani, J. Hoffman, and D. Batra (2022) ZSON: zero-shot object-goal navigation using multimodal goal embeddings. In Advances in Neural Information Processing Systems, External Links: 2206.12403, Document, Link Cited by: TABLE I, TABLE I. [21] D. Nie, X. Guo, Y. Duan, R. Zhang, and L. Chen (2025) WMNav: integrating vision-language models into world models for object goal navigation. arXiv preprint arXiv:2503.02247. Note: IROS 2025 External Links: 2503.02247, Document, Link Cited by: §1, §1, §2.1, §2.2, §3.2, TABLE I, TABLE I. [22] H. Niu, L. Zhang, X. Ji, R. Ying, P. Liu, and F. Wen (2026) SSMG-Nav: enhancing lifelong object navigation with semantic skeleton memory graph. Note: Accepted by ICRA 2026 External Links: 2603.01813, Document, Link Cited by: §1, §2.3. [23] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter (2019) Continual lifelong learning with neural networks: a review. Neural Networks 113, p. 54–71. External Links: Document Cited by: §2.3. [24] D. Qi, S. Zeng, X. Chang, F. Xiong, S. Xie, X. Wu, and M. Xu (2026) MerNav: a highly generalizable memory-execute-review framework for zero-shot object goal navigation. External Links: 2602.05467, Document, Link Cited by: §1, §2.1. [25] S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, M. Savva, Y. Zhao, and D. Batra (2021) Habitat-matterport 3d dataset (HM3D): 1000 large-scale 3d environments for embodied AI. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, External Links: Link Cited by: §4.1. [26] R. Ramrakhya, E. Undersander, D. Batra, and A. Das (2022) Habitat-web: learning embodied object-search strategies from human demonstrations at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5173–5183. Cited by: §2.1, TABLE I, TABLE I. [27] N. Savinov, A. Dosovitskiy, and V. Koltun (2018) Semi-parametric topological memory for navigation. In International Conference on Learning Representations, Cited by: §2.2. [28] M. Savva, A. Kadian, O. Maksymets, Y. Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V. Koltun, J. Malik, D. Parikh, and D. Batra (2019) Habitat: a platform for embodied AI research. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 9339–9347. External Links: Document, Link Cited by: §4.1. [29] X. Sun, L. Liu, H. Zhi, R. Qiu, and J. Liang (2024) Prioritized semantic learning for zero-shot instance navigation. In European Conference on Computer Vision, External Links: 2403.11650, Document, Link Cited by: TABLE I, TABLE I. [30] P. Wu, Y. Mu, B. Wu, Y. Hou, J. Ma, S. Zhang, and C. Liu (2024) VoroNav: voronoi-based zero-shot object navigation with large language model. External Links: 2401.02695, Document, Link Cited by: TABLE I. [31] K. Yadav, A. Majumdar, R. Ramrakhya, N. Yokoyama, A. Baevski, Z. Kira, O. Maksymets, and D. Batra (2023) OVRL-V2: a simple state-of-art baseline for ImageNav and ObjectNav. External Links: 2303.07798, Document, Link Cited by: TABLE I. [32] K. Yadav, R. Ramrakhya, A. Majumdar, V. Berges, S. Kuhar, D. Batra, A. Baevski, and O. Maksymets (2023) Offline visual representation learning for embodied navigation. In Workshop on Reincarnating Reinforcement Learning at ICLR, Cited by: §2.1, TABLE I. [33] H. Yin, X. Xu, Z. Wu, J. Zhou, and J. Lu (2024) SG-Nav: online 3d scene graph prompting for LLM-based zero-shot object navigation. Advances in Neural Information Processing Systems 37, p. 5285–5307. External Links: 2410.08189, Document, Link Cited by: §2.2. [34] N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher (2024) VLFM: vision-language frontier maps for zero-shot semantic navigation. In IEEE International Conference on Robotics and Automation, p. 42–48. External Links: 2312.03275, Document, Link Cited by: §1, §2.1, TABLE I, TABLE I. [35] B. Yu, H. Kasaei, and M. Cao (2023) L3MVN: leveraging large language models for visual target navigation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, External Links: Document, 2304.05501, Link Cited by: TABLE I. [36] M. Yu, F. Zhu, W. Liu, Y. Yang, Q. Wang, W. Wu, and J. Liu (2025) C-NAV: towards self-evolving continual object navigation in open world. In Advances in Neural Information Processing Systems, Note: NeurIPS 2025 External Links: 2510.20685, Document, Link Cited by: §1, §2.3. [37] L. Zhang, Y. Liu, Z. Zhang, M. Aghaei, Y. Hu, H. Gu, M. A. Alomrani, D. G. A. Bravo, R. Karimi, A. Hamidizadeh, H. Xu, G. Huang, Z. Zhang, T. Cao, W. Qiu, X. Quan, J. Hao, Y. Zhuang, and Y. Zhang (2025) Mem2Ego: empowering vision-language models with global-to-ego memory for long-horizon embodied navigation. External Links: 2502.14254, Document, Link Cited by: §1, §2.2. [38] S. Zhang, X. Yu, X. Song, X. Wang, and S. Jiang (2024) Imagine before go: self-supervised generative map for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, External Links: Link Cited by: TABLE I, TABLE I. [39] Q. Zheng, D. Liu, C. Wang, J. Zhang, D. Wang, and D. Tao (2024) ESceme: vision-and-language navigation with episodic scene memory. International Journal of Computer Vision. External Links: Document, 2303.01032, Link Cited by: §2.2. [40] L. Zhong, C. Gao, Z. Ding, Y. Liao, H. Ma, S. Zhang, X. Zhou, and S. Liu (2024) TopV-Nav: unlocking the top-view spatial reasoning potential of MLLM for zero-shot object navigation. arXiv preprint arXiv:2411.16425. External Links: 2411.16425, Document, Link Cited by: §2.2. [41] K. Zhou, K. Zheng, C. Pryor, Y. Shen, H. Jin, L. Getoor, and X. E. Wang (2023) ESC: exploration with soft commonsense constraints for zero-shot object navigation. In International Conference on Machine Learning, External Links: 2301.13166, Document, Link Cited by: TABLE I, TABLE I.