{"id":"2bbb821f-fa23-4e39-ae67-8ac6b6c10ecc","arxiv_id":"2608.10449","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PBD-AG is a persistent scene-graph system that separates verified stable fixtures from dynamically revised object events, and its simulation results show improved coarse F1, identity continuity, and event recall over adapted baselines.","lead":"PBD-AG builds a robot world model that keeps stable room fixtures in an unchangeable baseline and records object changes as typed events, using a camera-frustum visibility check before deciding an object is gone. In simulations it reports higher fixture and identity scores than adapted baseline systems, plus a physical demonstration.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's PBD-AG row appears to reuse Table 3's full-run MAE (0.356±0.233) despite being a restricted shared-evidence subset; the controlled comparison's 11.1/17.2-point advantage needs verification.","rationale":"The reader's weakest_assumption focused on the geometric visibility gate in Section 3.3, which is a plausible secondary concern about generalizability. However, the most load-bearing threat to the paper's central quantitative claim is the apparent internal inconsistency between Table 1 and Table 3. The shared-evidence comparison is the direct support for the headline 'exceeds DynaMem-adapted and ConceptGraphs-adapted coarse F1 by 11.1 and 17.2 points.' If the PBD-AG row in that comparison is not actually computed on the restricted subset, the claimed advantage is invalid or at least unverifiable from the manuscript. This is a concrete, checkable issue rather than a hypothetical failure mode. The reader's verdict of CONDITIONAL remains appropriate because the issue could be a reporting typo, but it must be resolved before the central comparison can be accepted. I therefore keep the verdict unchanged while flagging a concern that the reader's weakest_assumption did not identify.","tokens_in":9418,"tokens_out":14354,"duration_ms":133223,"concrete_test":"Obtain from the authors the per-seed evaluation records (or the evaluation harness) for Table 1 and Table 3, and recompute Table 1's PBD-AG row using only the shared category-intersection subset described in Section 4.1. If the resulting MAE and F1 differ from Table 3's All-rollouts row, the printed Table 1 is incorrect and the control comparison must be re-run; if they are identical, the subset restriction was not applied and the 11.1/17.2-point advantage over the adapted controls is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that the shared-evidence coarse comparison is 'reported separately from autonomous full-graph evaluation' and that 'the shared-evidence subset may therefore be smaller than the full-scene ground truth in Table 3.' Yet Table 1's PBD-AG row reports MAE = 0.356±0.233, identical to Table 3's 'All rollouts' MAE to three decimals, and F1 = 0.868±0.102 versus 0.866±0.102. The precision and recall values differ slightly, but an exact match of MAE and its standard deviation across nine seeds is not expected if the category subset actually changed. If the PBD-AG row was not recomputed on the restricted subset, then the comparison against DynaMem-adapted and ConceptGraphs-adapted is not apples-to-apples: the claimed 11.1-point and 17.2-point F1 advantages, and the count-MAE reduction from 0.716 to 0.356, would rest on numbers that do not correspond to the described protocol. Since no code or evaluation logs are released, the reader cannot distinguish a typo from an invalid comparison, and the central shared-evidence claim is not yet firmly supported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PBD-AG proposes a persistent baseline-delta scene graph for service robots: the robot bootstraps stable fixtures through frontier exploration and RGB-D observations, verifies them with close-range inspection, and maintains dynamic object states through evidence-gated persistent tracks with a geometric visibility gate to avoid false deletions under occlusion. The paper evaluates the approach in OmniGibson across three scenes with three seeds each (nine rollouts), reports coarse/fine F1 in Table 3, compares against DynaMem-adapted and ConceptGraphs-adapted controls under shared evidence in Table 1, and runs a controlled dynamic-memory protocol with 12 scheduled events across three Hotel sequences in Table 2, plus component ablations and a qualitative physical-robot demonstration.","tokens_in":9720,"tokens_out":9553,"duration_ms":87152,"significance":"If the reported numbers hold, the paper makes a useful empirical contribution: the shared-evidence comparisons in Tables 1 and 2 are thoughtfully designed, the component ablations isolate the visibility gate and persistent-ID contributions, and the fixed-parameterization claim across three environments is a practical strength. The dynamic protocol's 11/12 event recall and zero identity switches for PBD-AG are concrete, falsifiable results. The qualitative physical-robot demonstration supports the system-integration claim. The main gap is missing quantitative support for the active-inspection scheduler, which the manuscript asserts in the conclusions but does not actually present.","major_comments":[{"comment":"The manuscript repeatedly promises a third controlled protocol—paired independent active rollouts—and §5 asserts that 'Paired acquisition experiments further support graph-conditioned fixture prioritization under bounded budgets,' but no such experiment, table, or numeric result appears in §4 or in the provided text. Since active inspection is one of the four listed contributions, this is a load-bearing unsupported claim. The authors must either report the paired active-acquisition results (e.g., a table comparing PBD-AG's scheduler with random, frontier-only, or coverage baselines under a fixed inspection budget, with means and standard deviations over seeds) or explicitly remove this claim from the conclusions.","section":"§4.1, §5"},{"comment":"The headline results depend on a dozen hand-set parameters—the evidence increments (η_o, η_c, η_m, η_r) in Eq. (3), the lifecycle thresholds (0.72, 0.30, 0.12), the visibility-gate depth tolerance (0.15 m) and two-sample rule, the weights w in Eq. (4), and the cost coefficients in Eq. (6). The paper states that a single parameterization is held fixed across scenes, which is good, but it does not provide any sensitivity or perturbation analysis. Given that the claimed 0.014 false-absence rate and the 11/12 event recall are direct consequences of the visibility gate, a small sweep over the depth tolerance and existence-increment magnitudes is needed to show that the results are not an artifact of a narrow operating point.","section":"§3.3–§3.4, §4.2"}],"minor_comments":[{"comment":"Please state explicitly whether the PBD-AG row in Table 1 was recomputed on the shared-evidence subset and report the per-scene ground-truth counts of that subset. I do not see literal reuse of Table 3 (Table 3 contains no MAE column, and the P/R values differ), but the near-identical F1 mean and standard deviation (0.868±0.102 vs 0.866±0.102) will confuse readers without a clarifying sentence.","section":"§4.1, Table 1/Table 3"},{"comment":"The false-absence metric is computed 'over eligible post-acquisition checkpoints,' but the eligibility rule is never defined in the main text. Please specify the checkpoint-selection rule (time window, object state, visibility conditions) so the 0.014 false-absence result is reproducible.","section":"§4.1, Table 2"},{"comment":"The sentence 'Successful processing attaches verified child nodes and support edges to the existing canonical fixture and marks its close-range inspection complete' appears twice in Section 3.2; please remove the duplicate.","section":"§3.2"},{"comment":"Please clarify whether the three Hotel sequences in the dynamic protocol of Table 2 are the same runs as the Hotel rollouts in Table 3, and state whether event times were chosen a priori or post hoc.","section":"§4.3"},{"comment":"The manuscript mentions 'development and run-selection provenance, resolved configurations, and artifact hashes' in the supplementary material, but that material is not included in the arXiv text; if this provenance is part of the reproducibility claim, it should be made available on the project page.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Table 1 reusing Table 3's MAE does not land as stated: Table 3 has no MAE column, and the precision/recall values differ. The more serious issue is the missing paired active-acquisition results, which the conclusion asserts without evidence. This is fixable, and the remaining empirical design is credible for a robotics venue. I would not reject on the current evidence; I would require the authors to add the missing experiments or remove the claim, and to clarify the shared-evidence subset reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look. The baseline-delta split, the typed event audit stream, and the geometric visibility gate are real ideas, not just a re-packaging of existing scene-graph systems. The controlled evaluation design is better than what this subfield usually does: shared-evidence streams for the memory comparison, a separate dynamic benchmark with hidden event times, and ablations that isolate the visibility gate and persistent identity. The autonomous construction protocol over nine rollouts with fixed taxonomy and parameters is also a reasonable effort. Credit where it's due: Tables 1 and 2 are well designed and the headline claims about identity continuity (zero switches) and event recall (11/12) are clearly stated.\n\nSoft spots, in proportion. First, the active-inspection results are asserted in the conclusion ('paired acquisition experiments further support graph-conditioned prioritization') but no table or numbers appear anywhere. That's a real omission for a system whose title includes 'active graphs.' Second, no code, data, or evaluation logs are released; adapted baselines are explicitly not reproductions, so the strength of the 11.1- and 17.2-point F1 advantages is only as good as the common-interface implementation, which isn't auditable. Third, the hand-fixed thresholds in Equations 3 through 6 (η's, visibility depth window, promotion/demotion bounds) are numerous and no sensitivity analysis is provided. The visibility gate's 0.15 m depth tolerance and the two-sample rule are load-bearing for the low false-absence rate, and we just have to take them on faith.\n\nOn the stress-test concern: it doesn't hold up. Table 3 does not report MAE at all, so the 'identical MAE' claim is factually wrong. F1 values in Table 1 (0.868±0.102) and Table 3 (0.866±0.102) differ, as do precision and recall. So there's no evidence of an exact copy-paste. That said, the closeness is suspicious enough that I'd ask for the shared-evidence subset's size and the per-run numbers if I were refereeing — but the stress test's specific accusation is not supported.\n\nThis paper is for people working on long-horizon robot memory, scene graphs, and active perception. It deserves a serious referee, not a desk reject. My recommendation: send it to review, but insist the authors release code or at least detailed per-run logs, and add the missing active-inspection results. The central persistence claims are plausible and the architecture is genuinely interesting; it just needs to be checkable.","headline":"A genuinely new baseline-delta architecture for persistent robot scene graphs with well-designed controlled comparisons; the main issues are missing artifacts, an unshown active-inspection experiment, and hand-tuned thresholds — not the exact-match anomaly the stress test flagged.","tokens_in":10229,"tokens_out":2269,"would_cite":true,"duration_ms":21315,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a baseline–delta scene graph with visibility-gated negative evidence gives long-horizon robots persistent, identity-stable world models, reporting coarse F1 0.866±0.102 and event recall 11/12 in simulation.","keywords":["persistent world models","scene graphs","robot memory","visibility gating","active inspection","semantic mapping","long-horizon robotics","object identity"],"falsifier":"Replay the controlled dynamic benchmark while injecting a known localized camera-pose error of 5–10 cm during the disappearance event; if a still-present object is removed because its stored bounding box projects to the two required samples and the depth check passes despite the pose error, the visibility gate's depth tolerance is too tight for real localization noise. Concretely, compare the false-absence rate against the reported 0.014 under this injected error.","tokens_in":9239,"feed_emoji":"🤖","tokens_out":7999,"duration_ms":62464,"temperature":0.7,"pith_summary":"This paper argues that a long-horizon service robot can maintain a persistent world model by separating the scene into a stable structural baseline and a set of revisable object-event deltas, rather than continuously remapping or trusting each observation equally. The proposed PBD-AG system boots a fixture graph from scratch through exploration, verifies fixtures with close-view inspection, and then records later object changes as typed events attached to canonical identities. Its central move is a geometric visibility gate that allows a missed detection to count as evidence of absence only when the stored 3D bounding box was actually observable; otherwise the miss is ignored. Across nine simulated rollouts the paper reports coarse and fine F1 of 0.866±0.102 and 0.713±0.167, and in a controlled dynamic benchmark event recall of 11/12 with zero identity switches and a false-absence rate of 0.014. A qualitative physical-robot demonstration shows the full pipeline on onboard RGB-D sensing.","feed_headline":"Robot world model keeps object identities stable across 9 rollouts","feed_subtitle":"PBD-AG separates fixed fixtures from object changes and cuts false deletions under occlusion.","key_machinery":"The central object is the baseline–delta graph, a persistent runtime track table $M_t$ with a canonical-fixture view $eB_t$ and an ordered audit-event stream $E_{1:k_t}$. Each track holds class and existence beliefs, a 3D center with recent-position covariance, support-parent belief, an appearance embedding, and a persistent identity $h_i$ with a lifecycle state. The load-bearing mechanism is the geometric visibility gate in the existence update: unmatched tracks receive negative evidence only when $V_i(t)=1$, which requires at least two samples of the projected 3D bounding box to be in range, inside the image, and to pass $d_{\\mathrm{meas}} + 0.15\\,\\mathrm{m} \\geq d_{\\mathrm{expected}}$. This gate is what converts the system's claim about persistence into a concrete algorithm, because it decides when a missed detection counts as evidence that an object is gone.","core_discovery":"The central claim is that the occlusion-versus-absence problem—whether a missed detection means an object is gone or merely hidden—can be resolved by conditioning negative evidence on geometric observability, and that the right representation for persistence is a baseline–delta split: a frozen canonical-fixture view plus an ordered event stream. Concretely, PBD-AG maintains for each persistent track an existence log-odds updated by fixed increments for positive observation, cached confirmation, unmatched track, and cached rejection; the unmatched-track term is multiplied by a visibility indicator that is 1 only when at least two samples of the stored 3D box are in valid range, inside the image, and pass the depth-consistency check $d_{\\mathrm{meas}} + 0.15\\,\\mathrm{m} \\geq d_{\\mathrm{expected}}$. Under shared evidence the paper reports coarse F1 of 0.868±0.102, exceeding DynaMem-adapted and ConceptGraphs-adapted by 11.1 and 17.2 points, respectively, while in the dynamic protocol it achieves IDF1 0.833±0.029 with 11/12 event recall and zero identity switches. The ablation without the visibility gate raises false absence from 0.014 to 0.262 and cuts event recall from 11/12 to 6/12, which is the paper's main evidence that the gate is the load-bearing component.","pith_inferences":["One testable extension is calibrating the gate's depth tolerance (the fixed 0.15 m buffer) and the two-sample threshold; the paper's fixed choices imply an operating point on a curve trading event recall against false absence that is not explored.","The baseline–delta representation may transfer to other domains where absence evidence is unreliable, such as inventory or surveillance, wherever the core ambiguity is distinguishing 'gone' from 'hidden'.","Because identity association uses SigLIP cosine similarity above 0.78 with a unique-class fallback, identity stability is likely to degrade for appearance-similar object classes; the paper does not quantify that boundary.","The size of the improvement over adapted baselines suggests that the representation and memory-update rules, rather than detector quality or sensing, are the main drivers of persistence performance in these benchmarks."],"forward_implications":["Long-horizon robot deployments could maintain a compact, traceable world model without re-mapping the whole space after object changes; only the event suffix attached to the frozen baseline needs updating.","Planners can trust object identity and support relations across sessions, because the system reports zero identity switches in the controlled dynamic benchmark and an audit trail of typed events.","The visibility gate gives a principled rule for when repeated misses justify removal, which should reduce the classic failure where a robot deletes a present object because a detector or viewpoint failed.","Active inspection guided by graph uncertainty can concentrate bounded sensing and model-call budgets on unresolved fixtures rather than exhaustive re-scanning."],"supporting_citations":[{"why":"DynaMem; the adapted version is the stronger controlled baseline that PBD-AG must beat under shared evidence.","marker":"Liu et al. 2024"},{"why":"ConceptGraphs; supplies the other adapted baseline and the open-vocabulary scene-graph comparison.","marker":"Gu et al. 2024"},{"why":"SigLIP embeddings; provides the open-vocabulary feature evidence used in the persistent state.","marker":"Zhai et al. 2023"},{"why":"Grounding DINO; localizes selected fine-object phrases during inspection.","marker":"Liu et al. 2023"},{"why":"SAM 2; refines masks for fine proposals that back-project to 3D centers.","marker":"Ravi et al. 2024"},{"why":"BEHAVIOR-1K/OmniGibson; the simulation environment for the nine autonomous rollouts.","marker":"Li et al. 2024"},{"why":"IDF1 and identity-switch metrics; the dynamic evaluation uses these to score identity continuity.","marker":"Ristani et al. 2016"},{"why":"Frontier-based exploration; the coverage mechanism that expands unknown space during bootstrap.","marker":"Yamauchi 1997"}],"fun_headline_variants":["Robots get a persistent world model that separates stable from change","Baseline-delta graph with visibility gate cuts false deletions","PBD-AG: 11/12 event recall, zero identity switches in long-horizon","Occlusion-aware robot world model keeps object states truthful","Robot maps separate frozen fixtures from evolving object events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The low false-absence rate depends on the geometric visibility gate: a stored 3D bounding box is treated as observable when at least two samples are in range, inside the image, and satisfy measured depth plus 0.15 m at least expected depth; if this depth-window test mistakes occlusion or bounding-box drift for visibility, present objects receive negative evidence and can be demoted or removed.","fun_headline_variants_meta":{"raw":{"variants":["Robots get a persistent world model that separates stable from change","Baseline-delta graph with visibility gate cuts false deletions","PBD-AG: 11/12 event recall, zero identity switches in long-horizon","Occlusion-aware robot world model keeps object states truthful","Robot maps separate frozen fixtures from evolving object events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1610,"prompt_tokens":1079,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":695,"tokens_out":531,"duration_ms":5249,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:20:10.484629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replay the controlled dynamic benchmark while injecting a known localized camera-pose error of 5–10 cm during the disappearance event; if a still-present object is removed because its stored bounding box projects to the two required samples and the depth check passes despite the pose error, the visibility gate's depth tolerance is too tight for real localization noise. Concretely, compare the false-absence rate against the reported 0.014 under this injected error.","supporting_citations":[],"review_version":1}