{"id":"a545e876-49c9-49f5-811b-caf20e5e5ad2","arxiv_id":"2608.12707","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SAP-Nav combines an online spatial-semantic map with active viewpoint selection to let zero-shot agents navigate to targets specified by scene, room, region, and instance cues.","lead":"A new navigation system helps robots find objects from complex voice instructions like 'the black glass-top coffee table in the living room with the white fireplace' by actively choosing better camera views as they explore. It reportedly beats prior methods on two standard benchmarks, but the code and several key settings are not yet public.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline region-level advantage is confounded by VLM scale: with the same 7B backbone SAP-Nav beats PlaNaVid by only 2.0 points on Region and trails on other levels, so the 12.2-point claim does not isolate the active-perception contribution.","rationale":"The reader's strongest claim rests on the external SOTA comparison, and the most load-bearing condition is that the reported margin reflects the proposed framework rather than an uncontrolled difference in VLM size. Table 1 shows that with the same Qwen2.5-VL-7B backbone, SAP-Nav's region-level advantage over PlaNaVid is only 2.0 points, and SAP-Nav is worse than PlaNaVid at the other three granularities. The large 12.2-point margin appears only in the Qwen3-VL-235B row, so the paper's headline comparison is confounded. This is not an accusation of fabrication; it is a request for the obvious control condition. The internal ablations are genuine supporting evidence that QSSR and AVV contribute, but they do not resolve the external comparison. The reader's identified weakest assumption about AVV's geometric-visibility proxy is plausible but secondary: Table 4 shows a large overall gain from AVV, so even if the proxy is imperfect, the module demonstrably helps in these environments. The model-scale confound is more directly tied to the paper's central claim of state-of-the-art performance. I therefore keep the conditional verdict but add the matched-backbone comparison as an explicit condition for accepting the headline claim.","tokens_in":12792,"tokens_out":9222,"duration_ms":92088,"concrete_test":"Re-run the LangMap single-goal evaluation with PlaNaVid using the same Qwen3-VL-235B-A22B-thinking backbone, keeping the same detection, segmentation, and evaluation protocol as Table 1. Also report SAP-Nav with the same 235B model for direct comparison. If PlaNaVid-235B reaches or exceeds 43.8% region-level SR, the claimed 12.2-point advantage is a VLM-scale artifact and the central SOTA claim is unsupported; if PlaNaVid-235B remains near 31.6%, the active-perception modules are responsible for the margin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest external claim is the 12.2-point region-level margin (43.8% vs PlaNaVid's 31.6% in Table 1). This comparison is not controlled for VLM capacity: SAP-Nav's winning row uses Qwen3-VL-235B-A22B-thinking, while PlaNaVid is reported with Qwen2.5-VL-7B. The same table provides the needed control row: with the same Qwen2.5-VL-7B backbone, SAP-Nav reaches only 33.6% region SR, a 2.0-point margin, and it trails PlaNaVid on Single-Goal (29.4 vs 31.4), Scene (28.8 vs 34.4), Room (30.1 vs 35.6), and Instance (25.1 vs 26.2). Thus the 12.2-point advantage may be almost entirely attributable to the 235B VLM rather than to QSSR or AVV. The internal ablations in Tables 3 and 4 do show large gains from both modules, so the modules are probably doing something; but the paper's central claim of 'highest reported success rate' and '12.2% improvement over training-based methods' is not currently established as a property of the SAP-Nav framework rather than of its large backbone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SAP-Nav, a fully online, zero-shot framework for hierarchical open-vocabulary object navigation (OVON). It combines a Queryable Spatial-Semantic Representation (QSSR), built incrementally from online room segmentation and actively acquired room snapshots, with an Active Viewpoint Verification (AVV) module that repositions the agent when the current observation is judged insufficient for verifying a candidate target. The method is evaluated on the LangMap hierarchical benchmark and the HM3D-OVON standard benchmark, with additional qualitative real-world robot experiments. The headline result is a 43.8% region-level success rate on LangMap with a Qwen3-VL-235B-A22B backbone, reported as a 12.2-point improvement over the training-based PlaNaVid baseline, and a 49.7% success rate on HM3D-OVON with GPT-4o.","tokens_in":13129,"tokens_out":6550,"duration_ms":67142,"significance":"If the results hold, SAP-Nav would demonstrate that a zero-shot, fully online system can handle scene-, room-, region-, and instance-level goals without task-specific training or precomputed maps, outperforming training-based methods on hierarchical OVON. The paper's strengths are its broad benchmark coverage, internal ablations of QSSR and AVV, and a real-world deployment study. The central quantitative claim, however, is currently tied to the choice of a very large VLM; under a controlled same-backbone comparison the advantage over the training-based baseline is small and negative on most granularity levels. The active-perception design is a useful contribution, but the evidence that the framework itself, rather than the VLM scale, drives the headline gains requires a controlled comparison.","major_comments":[{"comment":"The abstract's headline claim of a 12.2% improvement in SR over training-based methods on region-level navigation is confounded by VLM scale. In Table 1, the winning SAP-Nav row uses Qwen3-VL-235B-A22B-thinking and reports 43.8% region SR, while PlaNaVid is reported with Qwen2.5-VL-7B at 31.6%. The same table provides the needed control row: with the same Qwen2.5-VL-7B backbone, SAP-Nav reaches 33.6% region SR, a 2.0-point margin, and it trails PlaNaVid on Single-Goal (29.4 vs 31.4), Scene (28.8 vs 34.4), Room (30.1 vs 35.6), and Instance (25.1 vs 26.2). The 12.2-point claim should therefore be presented as a property of the full configuration with a 235B VLM, or better, supported by re-running PlaNaVid with the same large VLM. Without this controlled comparison, the paper's central claim that SAP-Nav beats training-based methods is not established as a property of the framework.","section":"Abstract and Table 1"},{"comment":"No error bars, variances, or numbers of seeds are reported in Tables 1, 2, 3, or 4. Given the stochastic nature of VLM prompting, object detection, and exploration, differences of a few percentage points—such as the Instance-level 26.8 vs 26.2 in Table 1 or several ablation entries in Table 4—may be within run-to-run noise. The authors should report means and standard deviations over at least three seeds for the main configurations, and ideally for all compared methods.","section":"Experimental Setup and Tables 1-4"},{"comment":"The comparison across baselines is not fully controlled for the perception stack. Only VLFM is described as re-run with its official implementation; for the other baselines it is not stated whether their reported numbers come from the original papers, nor which object detector, segmenter, VLM, and frontier-exploration policy are shared. Since Table 2 includes methods also using GPT-4o (TANGO, MetaNav) that may differ in other components, the reader cannot separate the proposed framework's contribution from the chosen off-the-shelf components. The authors should specify the exact shared components for every baseline and mark each number as re-run or taken from the cited paper.","section":"Tables 1 and 2"},{"comment":"The viewpoint-selection module uses a geometric 2.5D height-map visibility score as a proxy for whether a new viewpoint will improve VLM verification, but the paper does not validate this proxy directly. The Limitations section acknowledges that AVV 'selects viewpoints based on geometric visibility, without considering their semantic utility or motion cost.' The Table 4 ablation shows that AVV improves performance, but it does not isolate whether the gain comes from the specific visibility model in Eq. (6) or simply from moving closer to the candidate and enlarging it in the image. An additional ablation that replaces Eq. (6) with a random navigable viewpoint near the candidate would help separate these explanations and support the transferability of AVV beyond these benchmarks.","section":"Method, Viewpoint Selection, Eq. (6)"}],"minor_comments":[{"comment":"The real-world deployment section is qualitative: it provides illustrative examples but no instruction-level success counts or quantitative task outcomes. Since the paper states that this 'validate[s] its practical feasibility,' the claim would be stronger with at least a small table of task-level outcomes across a fixed instruction set.","section":"Real-world Deployment"},{"comment":"In Eq. (3), the case C_k = {c} is well defined, but the behavior when C_k is empty or when the argmax over a multi-type set is not unique is unspecified. A sentence clarifying these edge cases would prevent ambiguity.","section":"Eq. (3)"},{"comment":"The claim of a '12.2% improvement' should specify that it is in absolute percentage points and that it is measured relative to a different VLM configuration, as discussed in the major comments.","section":"Abstract and text"},{"comment":"The caption says 'Single-Goal reports the average performance over all single-goal episodes,' which is vague about whether the average is over the four granularity levels or over a combined set of episodes. Please clarify the aggregation rule.","section":"Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"One of the paper's authors (B. Miao) is also a co-author of LangMap, the benchmark used for the headline evaluation. The benchmark is public and has external baselines, so this is not in itself a correctness problem, but the editor may wish to verify that the PlaNaVid and other LangMap comparison numbers are taken from the benchmark paper without cherry-picking. The paper also cites a large number of 2026 preprints; the novelty of combining online spatial-semantic memory with active viewpoint selection is nonetheless distinctive, and the controlled same-backbone comparison in Table 1 should be the basis for the paper's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe paper to know: SAP-Nav is a fully online, zero-shot system for hierarchical open-vocabulary object navigation, combining an online room-snapshot representation (QSSR) with an active-viewpoint verifier (AVV). It reports top results on LangMap and HM3D-OVON. The genuinely new part is treating observation acquisition as an explicit objective in this setting: QSSR actively captures holistic room views and AVV repositions before verifying candidates. Each block is adapted from known pieces, but the orchestration for multi-granularity goals is new, and the ablations in Tables 3 and 4 show both modules earn their keep. The paper is also honest enough to flag, in Limitations, that AVV picks viewpoints by geometric visibility, not semantic utility or motion cost.\n\nThe soft spot is the headline number. The abstract's 12.2% region-level gain compares SAP-Nav with Qwen3-VL-235B-A22B-thinking against PlaNaVid with Qwen2.5-VL-7B. The same table gives the control: with the same 7B backbone SAP-Nav's region margin shrinks to 2.0 points, and SAP-Nav trails PlaNaVid on Single-Goal, Scene, Room, and Instance. So the 12.2-point claim does not isolate the active-perception contribution; most of it is VLM scale. That is not fatal — the 9B row already beats PlaNaVid everywhere, and the ablations show real module gains — but the abstract and conclusion should say the margin depends on a larger VLM.\n\nOther concerns are smaller. There are no error bars or seeds, so I cannot tell whether the 2-point same-backbone region gap is noise. Several hyperparameters are unspecified (threshold tau, temporal window m, snapshot count Nv, sampling radii). Code is withheld until acceptance. The real-world section is qualitative. The LangMap benchmark is co-authored by one of the authors, but it is public and has many external baselines, so I do not read that as circular.\n\nWho is this for: people building zero-shot embodied navigation or VLM-based exploration. It deserves a serious referee. I would recommend conditional acceptance: require code release, per-seed variance, and a same-backbone headline comparison or a reworded claim.","headline":"Solid zero-shot active-perception system for hierarchical OVON, but the headline 12.2-point region gain is mostly VLM scale rather than the architecture; worth reviewing with revisions.","tokens_in":13666,"tokens_out":2607,"would_cite":true,"duration_ms":25431,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SAP-Nav claims a zero-shot, fully online navigation system can outperform training-based baselines on hierarchical open-vocabulary object navigation by actively choosing viewpoints and maintaining a queryable spatial-semantic…","keywords":["open-vocabulary object navigation","hierarchical goal navigation","zero-shot navigation","active perception","spatial semantic representation","vision-language model","viewpoint selection","online room segmentation"],"falsifier":"On a fixed set of LangMap region-level episodes, replace the Eq. (6) viewpoint selection with a randomly chosen navigable viewpoint from the same ring samples; if success rate does not drop meaningfully, geometric visibility is not carrying the reported improvement.","tokens_in":12618,"feed_emoji":"🤖","tokens_out":8676,"duration_ms":79745,"temperature":0.7,"pith_summary":"Open-vocabulary object navigation usually asks an agent to find any instance of a category, but people actually give goals that mix scene, room, region, and instance cues, such as \"the black glass-top coffee table in the living room with the white fireplace.\" SAP-Nav claims that these hierarchical goals can be solved fully online with no task-specific training and no precomputed map, by actively deciding where to look rather than passively using whatever the path happens to see. It builds a queryable spatial-semantic representation from room snapshots acquired on demand, and it repositions the robot before trusting a vision-language model's verdict on a candidate object. On the LangMap benchmark the best SAP-Nav configuration reaches 43.8% success on region-level goals, 12.2 points above the strongest training-based baseline, and on HM3D-OVON it reports the top success rate among compared methods.","feed_headline":"Zero-shot robot navigation beats trained baselines on hierarchical goals","feed_subtitle":"Online room semantics and viewpoint re-checking lift region-level success to 43.8% on LangMap.","key_machinery":"The load-bearing machinery is a pair of active-perception loops. QSSR fuses online room segmentation (door projections accumulated over a temporal window) with actively stitched room snapshots and a pretrained scene encoder's room-type logits, forming a BEV map where each cell carries room-type semantics and each room instance keeps a visual record. AVV scores the current view's sufficiency on visibility and perspective, samples navigable viewpoints on concentric rings, and selects the one maximizing the mean ray-cast visibility of the candidate footprint through an online 2.5D height map: $v^* = \\arg\\max_{v\\in\\mathcal{V}} \\frac{1}{|U(o)|}\\sum_{f\\in U(o)} \\mathrm{vis}(v,f)$. Only then does it ask the VLM to verify category and parsed attribute constraints, blacklisting rejected candidates.","core_discovery":"The paper's central claim is that active perception, applied at two distinct moments, is sufficient to close the gap that hierarchical open-vocabulary navigation opens: grounding room and region constraints needs persistent scene-level evidence, while verifying instance attributes needs a clear viewpoint, and no single passive trajectory reliably supplies both. To test this, SAP-Nav decomposes the problem into QSSR, which constructs an online BEV map whose cells carry room-type semantics and whose room instances carry stitched holistic snapshots, and AVV, which scores current-view sufficiency, selects a navigable and unoccluded viewpoint by ray-casting through an online 2.5D height map, and only then asks the VLM to verify category and parsed attributes. The reported results, best overall success rate on LangMap across scene, room, region, and instance granularities and the best success rate on HM3D-OVON, support the claim that a zero-shot system need not trade away multi-granularity competence for online operation.","pith_inferences":["The paper leaves implicit that the geometric visibility score in Eq. (6) could serve as a general \"should I look again\" trigger for other VLM-based perception tasks, such as instance re-identification or scene-graph verification, without retraining.","The per-episode QSSR limitation suggests a persistence extension: retaining room snapshots across tasks would let the same representation support lifelong navigation, with each visit accumulating evidence for later goal searches.","A clean stress test of the paper's core premise would randomize viewpoint selection while keeping everything else fixed; if success holds up, the gain would come from repositioning per se rather than from the geometric visibility model."],"forward_implications":["A fully online robot without a precomputed map can match or beat training-based navigators on goals that specify rooms, regions, and instances, not just object categories.","Room-level semantics can be extracted from a few actively captured holistic snapshots, so two rooms of the same type can be told apart during a single episode.","Repositioning before verification reduces false stops, and rejected candidates go to a blacklist so exploration resumes without re-querying the same object.","The same framework transfers to standard category-level object navigation, since scene-level success on HM3D-OVON improves without task-specific training."],"supporting_citations":[{"why":"Defines the LangMap benchmark and the scene-, room-, region-, and instance-level success metric that the paper's central comparison targets.","marker":"Miao et al. 2026"},{"why":"Supplies the HM3D-OVON dataset and benchmark used for the standard category-level object navigation evaluation.","marker":"Yokoyama et al. 2024b"},{"why":"Provides the vision-language frontier exploration backbone (VLFM) that SAP-Nav builds on for goal-driven exploration.","marker":"Yokoyama et al. 2024a"},{"why":"Contributes the online room segmentation method the paper adapts by accumulating door projections over a temporal window.","marker":"Bao et al. 2025"},{"why":"Provides the scene encoder whose room-type logits, restricted to an indoor vocabulary, form the sub-room semantic probability field.","marker":"Zhou et al. 2017"},{"why":"Documents the Qwen3-VL model family used for the top LangMap results, including the 235B-A22B configuration.","marker":"Bai et al. 2025a"},{"why":"Specifies the GPT-4o model used to obtain the top HM3D-OVON success rate reported in the paper.","marker":"OpenAI 2024"}],"fun_headline_variants":["SAP-Nav's active views beat trained baselines in zero-shot navigation","Zero-shot hierarchical OVON: active perception wins","Online spatial semantics and viewpoint checks top navigation","SAP-Nav: active perception for open-vocab navigation without training","See more, explore better: active perception for zero-shot navigation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing premise is that a viewpoint with higher geometric ray-cast visibility of the candidate's footprint in the online 2.5D height map will actually give the vision-language model enough visual detail to verify the object, because the whole AVV gain rests on that correlation.","fun_headline_variants_meta":{"raw":{"variants":["SAP-Nav's active views beat trained baselines in zero-shot navigation","Zero-shot hierarchical OVON: active perception wins","Online spatial semantics and viewpoint checks top navigation","SAP-Nav: active perception for open-vocab navigation without training","See more, explore better: active perception for zero-shot navigation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3357,"prompt_tokens":980,"completion_tokens":2377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2291}},"tokens_in":596,"tokens_out":2377,"duration_ms":17444,"temperature":1.0,"reasoning_tokens":2291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:07:22.544218+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed set of LangMap region-level episodes, replace the Eq. (6) viewpoint selection with a randomly chosen navigable viewpoint from the same ring samples; if success rate does not drop meaningfully, geometric visibility is not carrying the reported improvement.","supporting_citations":[{"cited_title":"2017 , publisher=","cited_arxiv_id":null,"evidence_quote":"Provides the scene encoder whose room-type logits, restricted to an indoor vocabulary, form the sub-room semantic probability field."}],"review_version":1}