acceptodds
Withdrawn from ICLR 2027

LANDER-Nav: Progressive Landmark-Schema Grounding for Multi-Granularity Open-Vocabulary Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Language-conditioned goal navigation is extending beyond category-level search toward multi-granularity goals disambiguated by landmarks, context, and spatial relations. Current methods encode instructions as a single joint query, conflating these constraints and often causing agents to bind the wrong same-category instance or stop at an unverifiable viewpoint. We disentangle each instruction into a landmark schema comprising target, anchor, context, and relation fields, and propose LANDER-Nav, a progressive grounding framework that preserves these role-specific constraints throughout navigation. Under this schema, LANDER-Nav coordinates three coupled decisions: Evidence Grounding decides where to explore by selecting frontiers likely to reveal missing evidence; Entity Grounding decides which entity to bind through landmark-anchored spatial voting over object memory; and Endpoint Grounding decides where to stop by choosing an executable viewpoint that verifies the required relation. Extensive experiments on LangMap and HM3D-OVON demonstrate that LANDER-Nav improves language-conditioned navigation, with particularly strong gains on long-horizon multi-goal tasks and context- and relation-dependent goals, while improving category-level open-vocabulary performance. Our code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Language-Conditioned Goal Navigation, Visual-Language Navigation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_abdaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LANDER-Nav, a modular framework for multi-granularity open-vocabulary navigation. The method decomposes language instructions into components such as target, anchor, context, and relation, and divides the navigation process into three stages: "Evidence Grounding" for selecting semantic frontiers; "Entity Grounding" for disambiguating target instances via multi-query retrieval and spatial voting; and "Endpoint Grounding" for selecting reachable viewpoints with visible targets.

The method was evaluated on the LangMap and HM3D-OVON datasets. Results demonstrate performance improvements across various settings; notably, LANDER-Nav outperforms MTU3D in fine-grained multi-target navigation tasks. Additionally, the paper presents component ablation studies and a comparative analysis based on different Vision-Language Model (VLM) backbones.

Critical strengths
  1. A navigation problem of practical significance driven by application needs.

This paper addresses a key challenge in language-conditioned navigation: how to leverage contextual landmarks and spatial relationships to identify the correct target instance, rather than relying solely on object categories. The distinctions drawn between exploration, instance disambiguation, and endpoint selection are sound, and the overall framework is clearly articulated.

  1. Improvements for multi-target navigation tasks involving detailed descriptions.

On the LangMap benchmark, LANDER-Nav outperformed MTU3D; notably, in multi-target navigation scenarios featuring detailed descriptions, the Success Rate (SR) rose from 41.8% to 45.5%, and the Success weighted by Path Length (SPL) increased from 25.1% to 28.6%. Although the magnitude of improvement varied across other evaluation settings, these results provide empirical support for the proposed method.

Critical weaknesses
  1. Limited technical novelty and insufficient support for progressive grounding.

The framework primarily combines VLM-based frontier reranking, multi-query object retrieval with spatial voting, and geometry-based endpoint selection, all of which build on established techniques. More importantly, the claimed explicit spatial relationship grounding is not clearly demonstrated. The spatial voting mechanism mainly relies on object proximity, without showing how directional relationships such as "left of," "right of," or "above" are explicitly modeled. Similarly, endpoint selection considers visibility, clearance, and reachability, but does not clearly verify target-anchor relationships. Consequently, the technical distinction between LANDER-Nav and existing structured or scene-graph-based navigation methods remains unclear.

  1. Limited performance gains and insufficient evidence for the full framework.

While LANDER-Nav improves performance on detailed multi-goal tasks, gains elsewhere are modest or inconsistent. Its SR on Single-goal Concise (28.6%) is below PlaNaVid (31.4%), and its improvements on Single-goal Detailed and HM3D-OVON are limited. Moreover, multi-goal experiments compare only against MTU3D. The ablation results also raise concerns about the contribution of the complete framework: Entity Grounding alone achieves 44.4% overall SR and 33.2% Instance SR, compared with 45.5% and 33.3% for the full method. Without leave-one-out ablations, the additional benefits of Evidence and Endpoint Grounding remain insufficiently established.

Questions for authors
  1. How are spatial relationships explicitly modeled and verified?

Spatial voting mechanisms rely primarily on geometric proximity, whereas endpoint grounding focuses on the visibility and reachability of the target. Could the authors clarify how different target-anchor spatial relationships are explicitly distinguished and constrained? In particular, how does the framework handle relationships such as "to the left of," "to the right of," or "above," which cannot be adequately characterized by proximity alone?

  1. What are the additional benefits of integrating all three grounding modules?

According to Table 3, when using only entity grounding, the overall Success Rate (SR) and Instance SR are 44.4% and 33.2%, respectively; for the complete framework, the corresponding figures are 45.5% and 33.3%. These differences are relatively small, particularly regarding instance targets. Could the authors clarify the complementary contributions of evidence grounding and endpoint grounding, and explain whether their integration yields significant benefits beyond using entity grounding alone?

Reviewer_c62MOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LANDER-Nav, a progressive landmark-schema grounding framework for multi-granularity open-vocabulary navigation. The framework addresses three key questions: which frontier to explore, which object to bind, and which viewpoint to select for stopping.

Critical strengths
  1. The paper addresses multi-granularity open-vocabulary navigation, where target objects are disambiguated by landmarks, context, and spatial relations. The proposed landmark schema explicitly represents target, anchor, context, and relation fields, providing a structured representation for context-dependent target disambiguation.
  2. The progressive grounding framework LANDER-Nav connects Evidence Grounding, Entity Grounding, and Endpoint Grounding to coordinate frontier selection, target binding, and stopping-viewpoint search. In particular, Endpoint Grounding considers target visibility, navigability, and obstacle clearance when selecting a reachable stopping viewpoint.
Critical weaknesses
  1. According to Table 1, LANDER-Nav underperforms PlaNaVid on several evaluation metrics across goal granularities, but the paper does not analyze these results. Moreover, why is the comparison with PlaNaVid limited to the single-goal concise setting, given that the LangMap paper introducing PlaNaVid formalizes the problem of multi-granularity open-vocabulary navigation? In addition, MTU3D, as the paper’s primary baseline, also outperforms LANDER-Nav on most metrics in the single-goal setting, yet the authors do not analyze this performance gap.

  2. The current ablation study does not adequately demonstrate the synergy among LANDER-Nav’s three components. In particular, Table 3 does not report results with all components disabled or with pairwise combinations of the components.

  3. The figures are not clearly annotated. In Figure 2, for example, the relationships among modules ①, ②, and ③, and how they fit into the overall pipeline, are unclear. Font sizes are also inconsistent. For example, the text in Figure 2 is too small to read easily, while the text in Figure 4 is too large.

Questions for authors

Please see the weaknesses.

1 comment

Given that the authors have not responded, I will maintain my original score.

Reviewer_WoFFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi-granularity open-vocabulary navigation, where a goal is specified by a target together with landmarks, room context, and spatial relations. The authors argue that existing methods compress such instructions into a single query, which leads to binding the wrong same-category instance or stopping at a poor viewpoint. LANDER-Nav parses each instruction into a target/anchor/context/relation schema and adds three gated modules on top of an MTU3D-style backbone: VLM-based frontier re-ranking, landmark-anchored spatial voting for target binding, and visibility-based stopping viewpoint search. The clearest gain is on LangMap detailed multi-goal navigation (45.5 vs. 41.8 SR over MTU3D), with smaller gains in other settings and +1.4 SR on HM3D-OVON.

Critical strengths
  1. The motivation is clear. LangMap-style goals do require instance disambiguation, and the paper correctly identifies that ranking candidates by a single sentence-level similarity score is a real limitation of existing methods.

  2. The target-binding module (Entity Grounding) is well supported by the ablation. Enabling it alone lifts Instance SR from 27.8 to 33.2 and Region SR from 38.1 to 44.0, while the other two modules barely change these levels.

  3. The design is practical. Each module falls back to the backbone when its gate fails, no extra training is needed, and the appendix gives the full decision procedure and prompts.

  4. The qualitative examples are informative, and performance is stable across VLMs. Fig. 4 directly shows MTU3D stopping at the wrong same-category object, and four different VLMs give similar results.

Critical weaknesses
  1. The paper claims that a unified landmark schema guides exploration, binding, and stopping, but in the implementation only Entity Grounding actually uses the decomposed schema. Frontier scoring sends the raw instruction to the VLM, the endpoint search only looks at target visibility without any anchor, and spatial relations like "above" or "beside" are reduced to simple distance. As implemented, the three modules do not exchange schema information, so the claimed "progressive" grounding is not supported by the method. Either such a mechanism should be added and evaluated, or the abstract and introduction should describe the method more accurately.

  2. Scene goals have no anchors but still improve from 52.6 to 55.5 SR, mostly due to the stopping-viewpoint module (Endpoint Grounding), and the HM3D-OVON gain is also category-level. So it is hard to tell how much of the main improvement comes from schema grounding rather than better stopping-viewpoint selection.

  3. In the appendix, switching from FastSAM to SAM 2.1 alone gives +3.6 SR, which is about the same as the +3.7 gain over MTU3D. The paper does not say which segmenter the MTU3D baseline uses, so an MTU3D + SAM 2.1 comparison is needed.

  4. The ablations do not isolate the contribution of the proposed components. A full-instruction VLM re-ranking baseline without role decomposition, and target binding without spatial voting, would directly test the two main claims. UniGoal and MSGNav, which the paper itself describes as the closest structured methods, are also not compared.

  5. Since the paper is motivated by wrong-instance binding and bad stopping, it would help to see how often these failures occur for MTU3D and LANDER-Nav, and how often each gate overrides the backbone correctly.

Questions for authors
  1. Is the backbone exactly the released MTU3D, and does the MTU3D row in Tables 1 and 2 use the same perception modules (segmenter, DINOv2, CLIP) with only the three modules turned off?

  2. In the single-goal concise setting, LANDER-Nav is 3.2 points below MTU3D on Region SR (29.5 vs. 32.7) but has the best Region SPL (17.3), which gives an SPL/SR ratio of 0.59, compared with 0.50 for MTU3D and 0.54 for LANDER-Nav itself under detailed descriptions. Is this MTU3D row re-run under your setup or taken from the LangMap paper, how often do the gates override the backbone in this setting, and could you report the path efficiency of successful Region episodes for both methods?

  3. How often do LangMap Instance goals contain relations that distance alone cannot resolve? Would a direction-aware support term (e.g., for "above" or "on") help?

  4. Qwen-VL-Plus outperforms newer models such as Qwen3-VL-8B and Gemini-2.5-Flash in Table 4. Were the prompts or thresholds tuned with Qwen-VL-Plus, and what does a no-VLM row on the same 100-subgoal subset give?

  5. The decomposition prompt refers to "RefHM3D" instructions. Where were the gate thresholds tuned, and were the LangMap evaluation scenes used for tuning? Their actual values are also not reported.

Reviewer_ymJcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses LangMap-style multi-granularity navigation, where the target instance must be disambiguated by landmarks, context, and spatial relations. LANDER-Nav is a training-free add-on to a 3D grounding-and-exploration backbone (MTU3D). A VLM splits the instruction into target, anchor, context, and relation fields, and three gated modules can override the backbone: VLM-based frontier re-ranking, anchor-supported spatial voting for target binding, and geometric stopping-viewpoint selection. The main gain is on LangMap detailed multi-goal navigation (45.5 vs. 41.8 SR over MTU3D); gains elsewhere are about one point or less, and HM3D-OVON SR improves from 46.9 to 48.3.

Critical strengths

The method targets a real failure mode, binding the wrong same-category instance, and its fallback gates make it a safe plug-in for existing policies. The clearest support is on detailed multi-goal Instance goals (33.3 vs. 27.8 SR), where Entity Grounding alone accounts for most of the gain.

Critical weaknesses

The implementation does not match the central claim. The paper states that the schema is preserved across all three decisions and that the stopping viewpoint verifies the required relation. In the method, however, the relation and room fields never enter any score, the spatial term reduces every relation to proximity, Evidence Grounding uses the raw instruction, and Endpoint Grounding uses only target geometry and occlusion. “Landmark-Ray Endpoint Search” appears only in the introduction.

Questions for authors

Can you add schema-free controls with a matched VLM budget: frontier re-ranking with the holistic instruction, and object re-ranking with the full-instruction query only (no anchor queries or spatial voting)? Please also add a backbone row and leave-one-out rows to Table 3. This determines whether the decomposition, rather than added VLM compute, drives the gains.

Are relations or room context used anywhere beyond the full-instruction query, and does Endpoint Grounding use the anchor at all? If not, the abstract, introduction and Fig. 1 should be brought in line with the method. If so, results broken down by relation type would support the relation claims.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.