{"id":"b7b8728f-7a93-47da-86ec-4445f5f60d07","arxiv_id":"2607.10437","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Interleaved high-level LLM-histogram POUCT and low-level particle-belief motion planning reduces collisions up to 63%, steps 35%, and detections 32% on multi-object search in unknown multi-room settings.","lead":"Inter-POMDP is a two-level planner that searches for multiple objects in unknown multi-room homes by letting a high-level semantic POMDP talk to a low-level obstacle-aware motion planner. It cuts collisions, travel, and camera uses versus prior hierarchical methods, which matters for practical household robots that must handle clutter and partial maps.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Headline gains rest on perfect detection and known furniture; Table I does not isolate the interleaving contribution under those assumptions.","rationale":"The reader correctly flags the perfect-observation / known-furniture assumptions as the weakest modeling premise. That premise is not merely a realism caveat; it is load-bearing for the causal claim that interleaving is what produces the Table I numbers. Because the paper supplies no ablation that isolates the upward cost-propagation step, and because both baselines also operate under the same idealized observation model, the quantitative support for the algorithmic novelty remains conditional on those assumptions. The recommended test (noisy detection + feedback ablation) would settle whether the gains survive once the high-level beliefs are no longer nearly deterministic. No internal contradiction is present, so the verdict stays CONDITIONAL rather than moving to REJECT; the concern simply sharpens the condition that must be met before the strongest claim can be accepted at face value.","tokens_in":11715,"tokens_out":560,"duration_ms":8407,"concrete_test":"Re-run the three ProcTHOR scenes of Table I with a single controlled change: inject realistic detection noise (e.g., YOLO-style false-negative rate 0.15–0.25 and occasional false positives inside the FOV) while keeping all other components fixed, and report the same three metrics for Inter-POMDP versus an ablated variant that disables the Q_init cost feedback. If the relative gains collapse or reverse under noisy detection, the headline claim does not hold outside the perfect-observation regime.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim attributes the reported gains (up to 63%/35%/32% fewer collisions/steps/detections vs CSG-TL and COSPOMDP) to the interleaved cost-feedback design. That attribution is load-bearing on the modeling choices in Section III: the observation model H is perfect (fan-shaped FOV of depth D, perfect accuracy; detection declared out of scope) and furniture is known a priori via Mo. Under those conditions the high-level histogram beliefs and LLM relations become nearly deterministic once a landmark is seen, so the residual uncertainty that the low-level particle beliefs and Q_init feedback are meant to resolve is mainly floor-obstacle geometry. Table I and Fig. 3 therefore compare full systems that also differ in map representations, LLM usage, and belief type; they do not ablate the upward cost-propagation step itself (Eqs. 3–4 and Algorithm 1 lines 10–11). Consequently it is unclear how much of the measured improvement is produced by interleaving versus by the perfect-observation + known-furniture regime that both the method and the baselines inherit.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Inter-POMDP, a two-level interleaved POMDP planner for multi-object search in unknown multi-room households. A high-level POUCT planner maintains LLM-informed histogram beliefs over object–furniture and object–object relations on a semantic map and selects abstract subtasks (Navigate/Search/Sense/Declare). A low-level motion planner maintains particle beliefs over unknown floor obstacles, builds a topological map with A*-derived edge costs, and feeds those navigation costs upward as domain knowledge that initializes Q-values and biases rollouts in the high-level UCT (Eqs. 3–4, Algorithm 1). The authors claim this interleaving yields up to 63% fewer collisions, 35% fewer navigation steps, and 32% fewer detections than CSG-TL and COSPOMDP on three ProcTHOR scenes (15 trials total) plus a qualitative real-robot breakfast-item search demo.","tokens_in":12079,"tokens_out":1684,"duration_ms":33300,"significance":"If the interleaving mechanism is the true source of the reported gains, the work would be a useful contribution to hierarchical POMDP planning for household object search: it extends prior sequential or non-interacting multi-scale planners by a concrete upward cost-feedback channel, combines LLM semantic priors with geometric particle beliefs, and includes both simulation and a real mobile-manipulator deployment. The dual-map design (semantic + topological) and the explicit treatment of floor obstacles as both hazards and (sometimes) targets are practically relevant. Strengths that should be credited include a clear algorithmic sketch (Algorithm 1), public video evidence, and head-to-head comparison against two recent baselines on the same ProcTHOR scenes. The significance of the central claim, however, currently rests on system-level comparisons that do not isolate interleaving, so the incremental value of the novel interaction mechanism remains only partially established.","major_comments":[{"comment":"The central claim attributes the Table I / Fig. 3 gains to the interleaved cost-feedback design (Algorithm 1 lines 10–11; Eqs. 3–4, Q(s,a)←Q_init from low-level costs). Table I and Fig. 3 compare full systems that also differ in map representations, LLM usage, belief type (histogram vs particle), and correlation models. There is no ablation that disables only the upward propagation of navigation costs (e.g., high-level POUCT with the same LLM priors and maps but Q_init=0 / no cost feedback, or sequential hierarchical control as in COSPOMDP). Without that control, it is not possible to credit interleaving—as opposed to particle obstacle beliefs, LLM priors, or topological mapping—for the reported reductions. An ablation (or at least a sequential hierarchical variant of Inter-POMDP) is load-bearing for the novelty claim.","section":null},{"comment":"Section III states that the observation model H is a fan-shaped camera of depth D with perfect accuracy and that object detection is out of scope; furniture is known a priori via the occupancy grid Mo. Under these assumptions, once a landmark is observed the high-level histogram belief conditioned on Coo/Cof becomes nearly deterministic, so residual uncertainty that the low-level particles and Q_init feedback resolve is mainly floor-obstacle geometry. The abstract and introduction frame the problem as extensive uncertainty over object locations and cluttered unobserved obstacles; the experimental regime substantially reduces the object-location component. The paper should either (i) relax perfect detection / known furniture in experiments, or (ii) explicitly scope the claim to obstacle-aware navigation under known furniture and perfect sensing, and re-evaluate how much of the multi-objec","section":null},{"comment":"Table I reports mean±std over only 5 trials per mission with no statistical tests (paired t-tests, bootstrap CIs, or multiple-comparison correction across 3 scenes × 3 objects × 3 metrics). Several cells show Inter-POMDP equal to or worse than a baseline on early objects (e.g., scene 8 object 1 steps: 229±3 vs CSG-TL 223±4; scene 13 object 1: 77±3 vs COSPOMDP 67±5; scene 1 object 2: 34±2 vs COSPOMDP 27±2), with gains concentrated on later objects. The abstract’s “up to 35%” navigation-step reduction is also hard to reconcile with the Analysis section’s “over 90% and 80%” claims for train-13 object 3. Aggregate totals and significance tests should be reported so that the headline percentages are reproducible from Table I and not driven by single late-object cells.","section":null},{"comment":"The mapping from low-level navigation costs c (Algorithm 1 line 8–9) into the high-level Q_init(s,a) (Eq. 4) is described only qualitatively (“encodes accumulated navigation costs”). No formula, normalization, or horizon scaling is given, nor is sensitivity to the UCT constant c or particle count reported. Because Q_init is the sole mathematical embodiment of interleaving, its definition must be precise enough for reimplementation and for verifying that asymptotic UCT guarantees are preserved under the chosen initialization.","section":null}],"minor_comments":[{"comment":"Real-world experiments (Fig. 5, §V.D) are purely qualitative; no collision/step/detection numbers are given for the physical platform, so the abstract’s joint claim of “simulation and real-world experiments” overstates the quantitative support from the robot demo.","section":null},{"comment":"Baselines CSG-TL and COSPOMDP are reimplemented under the same perfect-detection / known-furniture regime; the paper should state any adaptations required (e.g., how CSG-TL’s off-the-shelf navigation was replaced) so that differences are not confounded by implementation choices.","section":null},{"comment":"LLM prompt design is only illustrated in Fig. 2; the exact structured prompt, temperature, and how Mm memory updates are written back into subsequent prompts should be specified or released for reproducibility.","section":null},{"comment":"GAT relation-model training (300 epochs, batch 16, 1e-4, 1500 ProcTHOR pairs) is mentioned in §V.A but it is unclear whether the learned model is used at test time or only LLM relations (Eq. 1). Clarify the role of the GAT versus GPT-4o.","section":null},{"comment":"Typographical / presentation: “mutilple rooms” in Table I caption; “Full videos arehttps://…” missing space in abstract and table; arXiv-style line breaks leave hyphenated words mid-sentence throughout. Fig. 3 aggregated bars would benefit from error bars matching Table I stds.","section":null},{"comment":"Related work correctly positions against COSPOMDP [5] and OO-POMCP [7]; a short explicit contrast with DESPOT-style regularization and with hierarchical constrained MCTS [24] on the interleaving (vs. purely hierarchical) axis would help readers place the contribution.","section":null}],"recommendation":"major_revision","confidential_remarks":"The methodological idea is publishable with a proper ablation and tightened experimental claims; I would not reject on novelty grounds. The main risk is that without isolating interleaving the paper reads as a well-engineered system paper rather than a clear algorithmic advance. If the authors cannot run ablations, a major revision that rewrites the claim around the full system (and softens “first interleaved” attribution language) could still be acceptable. Scope fits a robotics / planning venue; the perfect-detection assumption is common in this literature but should be owned more carefully given the paper’s uncertainty framing."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit here is the interleaving itself: high-level POUCT with LLM-conditioned histogram beliefs over object-furniture relations, low-level particle beliefs over floor obstacles that return navigation costs as Q_init into the high-level UCT (Eqs. 3–4 and Algorithm 1). That feedback loop is not in COSPOMDP (sequential, no cost return) or the earlier MDP interleaving paper they cite. The rest of the stack—semantic + topological maps, LLM relation prompts, particle reinvigoration—is competent assembly of known pieces.\n\nWhat they do well is keep the two levels talking and show the system working end-to-end. Table I and Fig. 3 give consistent mean reductions across three ProcTHOR scenes (5 trials each) and a real dual-arm mobile base demo; collisions drop to zero after the first object, later objects take fewer steps and fewer camera activations. The citation pattern is honest about the hierarchical baselines and the MDP precursor. No internal contradiction; the math is standard UCT + particle filtering.\n\nSoft spots are real but not fatal. The observation model is perfect (fan FOV, perfect accuracy; detection declared out of scope) and furniture is known a priori via Mo, so residual uncertainty is mostly floor obstacles. Under those conditions the high-level beliefs become nearly deterministic once a landmark is seen, which makes the low-level cost feedback look more powerful than it may be in messier sensing. Table I compares full systems that also differ in map type, LLM use, and belief representation; there is no ablation that isolates the upward cost-propagation step. N is small, no statistical tests, free parameters (UCT c, particle count, GAT training) are listed but not swept. The “up to 63/35/32 %” language is the usual cherry-pick of best-case deltas.\n\nThis is for people already building hierarchical POMDPs for household search who want a concrete way to push geometric cost back into semantic action selection. It is not a theory paper and does not reorganize the field. I would send it to peer review; a referee can demand the ablation and a noisy-detection experiment. Worth a look if you are in that subfield; I would cite the interleaving mechanism if I were writing a related planner.","headline":"Solid hierarchical-POMDP engineering with a real interleaving idea; the headline percentages over-claim relative to the missing ablation and the perfect-detection setup.","tokens_in":12641,"tokens_out":554,"would_cite":true,"duration_ms":7089,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Interleaved high-level semantic and low-level geometric POMDP planners let a robot find multiple objects in unknown multi-room homes with far fewer collisions and steps.","keywords":["POMDP planning","multi-object search","hierarchical planning","interleaved planners","particle beliefs","LLM priors","household robotics","navigation under uncertainty"],"falsifier":"Run the same three-object search tasks in ProcTHOR scenes train_1, train_8 and train_13 with a deliberately noisy detector (or without the perfect-FOV assumption) and check whether the reported collision, step and detection reductions versus CSG-TL and COSPOMDP still hold.","tokens_in":12654,"feed_emoji":"🤖","tokens_out":662,"duration_ms":6874,"temperature":0.7,"pith_summary":"Household robots must search for several objects whose locations are unknown, while also navigating rooms full of unobserved obstacles. Full POMDP planners that treat the whole problem at once become intractable as the number of objects and rooms grows. This paper introduces Inter-POMDP, which splits the problem into two cooperating planners: a high-level POUCT planner that maintains histogram beliefs over furniture-level object locations, seeded by language-model priors about typical object–furniture relations, and a low-level particle-belief planner that reasons about obstacle-aware trajectories. The low-level planner continually returns measured navigation costs as domain knowledge that biases the high-level UCT action values and rollouts. The interaction lets the robot prune expensive or blocked search branches early, so that later objects are found with almost no extra collisions and far fewer camera detections. Experiments in three large simulated multi-room scenes and on a real dual-arm mobile robot show reductions of up to 63 percent in collisions, 35 percent in navigation steps and 32 percent in detections relative to two recent hierarchical baselines.","feed_headline":"Robot finds multiple objects with 63% fewer collisions","feed_subtitle":"Interleaved semantic and geometric POMDP planners cut steps and camera uses in unknown multi-room homes","key_machinery":"Interleaved domain-knowledge-augmented UCT: navigation costs computed by a particle-belief low-level planner are used both to initialize Q-values and to bias rollouts inside a high-level POUCT tree that reasons over LLM-informed histogram beliefs on a semantic map.","core_discovery":"An interleaved two-level POMDP architecture in which low-level obstacle-aware navigation costs are fed back as domain knowledge to initialize and bias high-level UCT selection produces near-optimal multi-object search policies that are both safer and more efficient than purely sequential hierarchical POMDPs or language-model scene-graph methods in unknown multi-room households.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Interleaved POMDP cuts multi-object search collisions by 63%","Two-level planner slashes household nav steps by 35%","Inter-POMDP reduces detections 32% in unknown multi-room homes","Obstacle-aware low-level beliefs guide high-level object search","Interleaved POMDPs yield safer multi-object policies in homes"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Object detection is assumed perfect inside a fixed camera field of view, and the locations of all furniture pieces are known in advance from the occupancy grid; the only unknowns are targets, landmarks and floor obstacles.","fun_headline_variants_meta":{"raw":{"variants":["Interleaved POMDP cuts multi-object search collisions by 63%","Two-level planner slashes household nav steps by 35%","Inter-POMDP reduces detections 32% in unknown multi-room homes","Obstacle-aware low-level beliefs guide high-level object search","Interleaved POMDPs yield safer multi-object policies in homes"]},"model":"grok-4.5","effort":"low","cost_usd":0.005032,"raw_usage":{"total_tokens":1384,"prompt_tokens":722,"num_sources_used":0,"completion_tokens":95,"cost_in_usd_ticks":50320000,"prompt_tokens_details":{"text_tokens":722,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":567,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":722,"tokens_out":95,"duration_ms":5048,"temperature":1.0,"reasoning_tokens":567,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T11:43:58.381360+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same three-object search tasks in ProcTHOR scenes train_1, train_8 and train_13 with a deliberately noisy detector (or without the perfect-FOV assumption) and check whether the reported collision, step and detection reductions versus CSG-TL and COSPOMDP still hold.","supporting_citations":[],"review_version":1}