{"id":"a76e97a5-cd02-45f4-aaeb-201873cccf42","arxiv_id":"2411.11733","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Integrating planner feedback into active sensing improves success rates for object retrieval from unknown confined cluttered spaces, reaching 95% success in 100 simulated scenarios.","lead":"A robot arm with an in-hand camera uses an integrated active sensing and Monte-Carlo Tree Search planner to retrieve target objects buried in unknown, cluttered, confined spaces. The system beats several baselines in simulation and one real cabinet test by letting the planner tell the sensor which unobserved regions matter most.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 95% success rate is a planning-generation metric (Sec. IV-A), not end-to-end retrieval; the one detailed real-robot success and acknowledged low-light segmentation failures leave execution transfer unsupported.","rationale":"Good faith reading: the paper is a systems paper, and its contribution is the bidirectional feedback loop. The ablations (IAS+SAS vs IAS+FAS vs MAS) do suggest that feedback from the planner to sensing matters, and the real-world video is supporting evidence. The concern is not that the simulation protocol is hidden or misleading—it is explicitly stated—but that the protocol measures plan generation, not the object retrieval task named in the title and conclusion. This is load-bearing because the abstract and conclusion claim effectiveness at retrieval; the comparison against baselines supports a narrower claim of planning efficiency. The absence of statistical tests and the omission of the most relevant baseline [20] further weaken external support, but these are secondary to the planning-vs-execution gap. A physics-execution check is the minimal experiment that would settle whether the integrated planner's feasibility checks survive contact with dynamics and perception noise; the reader's conditional verdict should therefore stand unchanged.","tokens_in":11061,"tokens_out":6739,"duration_ms":72931,"concrete_test":"Re-run the same 100 scenarios in a physics simulator (e.g., PyBullet) with the UR5e model, Contact-GraspNet grasps, and the planned pick-and-place trajectories executed with closed-loop control; score success only when the target object is physically moved out of the confined space. Compare this execution success rate with the reported 95% planning success. If the execution rate drops substantially (e.g., below 80%) or the ablation ranking changes, the central claim should be downgraded to a planning-only result pending more real-world trials.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central evaluation defines success as 'generate a feasible object retrieval plan before the scene is entirely observed' (Sec. IV-A). Table I/II therefore report planning success, not physical retrieval: no simulated plan is executed, and grasp feasibility, closed-loop control, and pick-place execution errors are not scored. For the conclusion that MAS+OR-MCTS 'solves object retrieval problems in confined, cluttered environments' (Sec. V), the unstated load-bearing assumption is that the geometric feasibility checks used by the planner transfer to real execution. That assumption is supported by only one detailed real-world demonstration among five scenes (Fig. 2), and the paper itself reports that the segmentation algorithm can fail in cluttered, low-lighting conditions (Sec. IV-C). Since the 95% number is computed on randomly generated simulated scenes with no physics or control layer, it cannot by itself establish that the integrated system retrieves objects reliably; it only establishes that the planner can find collision-free rearrangement sequences under an optimistic geometric model. A failure of this transfer would not necessarily overturn the relative ranking of planners in simulation, but it would invalidate the paper's central claim of efficient object retrieval in the real world.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an integrated system, MAS+OR-MCTS, for retrieving target objects from unknown, confined, cluttered environments. MAS combines three active-sensing stages—initial target detection, sensing of robot swept volumes, and feedback-based sensing from a failed plan—while OR-MCTS performs rearrangement planning with MCTS. The authors evaluate the approach in 100 randomly generated simulated scenarios and five real cabinet scenes, reporting a 95% planning-level success rate for the full system and favorable comparisons against ablations and three planner baselines. The paper claims that bidirectional feedback between active sensing and rearrangement planning is the decisive factor for efficient object retrieval.","tokens_in":11310,"tokens_out":5607,"duration_ms":56517,"significance":"If the empirical claims hold, the paper makes a useful contribution by demonstrating that coupling task-driven active sensing with rearrangement planning can substantially improve planning-level success in partially observed confined spaces. The evaluation design has strengths: a 100-scenario simulation suite, a reasonable ablation set (MAS, IAS+SAS, IAS+FAS, IAS), external planner baselines (SS-MCTS, TSAD, WTL), and a real-robot demonstration. The central limitation is that success is measured as generating a feasible plan, not executing it, and the real-world evidence is limited to one detailed instance. The contribution is plausible and potentially valuable, but the strength of the current claims exceeds the evidence.","major_comments":[{"comment":"The success metric in Section IV-A defines success as 'generate a feasible object retrieval plan before the scene is entirely observed,' not as physically executing that plan and retrieving the target. Consequently, the 95% success rate in Tables I and II is a planning-feasibility rate under a geometric model: simulated plans are not executed, and grasp success, closed-loop tracking, and pick-and-place errors are not scored. The only evidence for execution transfer is one detailed real-world instance in Fig. 2, and Section IV-C itself reports segmentation failures in cluttered, low-light conditions. The Section V claim that the method 'solves object retrieval problems' is therefore stronger than the evidence supports; either report end-to-end execution success rates in simulation and across all five real scenes, or restrict the claims to planning-level success.","section":"IV-A and IV-C"},{"comment":"The paper repeatedly states that the proposed method outperforms baselines 'by a significant margin,' but no statistical test, confidence interval, or paired comparison is reported. With N=100 scenarios, large mean differences in success rate (e.g., 95% vs. 92% vs. 61%) may or may not be statistically meaningful, especially for smaller differences such as MAS vs. IAS+FAS in planning time (10.20 ± 9.71 vs. 9.71 ± 9.57). Please report p-values or bootstrap confidence intervals for all pairwise comparisons, or remove the word 'significant.'","section":"IV-A, Tables I-II"},{"comment":"The comparison against 'state-of-the-art' methods omits reference [20], Ren and Qureshi (2024), which addresses exactly the same problem—object retrieval from confined spaces with a robot's in-hand RGB-D sensor—and is from the same group. Because [20] is cited in the related work but not included in Table II, the claim of outperforming state-of-the-art baselines is incomplete. Add this method as a baseline or justify its exclusion in the text.","section":"IV-B"},{"comment":"OR-MCTS is described only in prose: there is no pseudo-code, no definition of the tree policy beyond 'tuned Upper Confidence Bound,' and no values for the UCB exploration constant, the 30-second time budget distribution, the number of expansions or rollouts, or the stopping criterion for node evaluation. These details are required to reproduce the method and to assess whether the reported planning times depend on favorable tuning; please provide a full algorithm listing and a table of hyperparameters.","section":"III-B-2 and Algorithm 1"},{"comment":"The real-robot section reports only one detailed instance (Fig. 2) and says, without quantitative support, that results are 'similar in statistics' to simulation. No table lists per-scene success, planning time, number of attempts, or the segmentation failures mentioned in the final paragraph. Please provide a full table for the five scenes and report the failure count and failure modes.","section":"IV-C"}],"minor_comments":[{"comment":"Line 6 of Algorithm 1 uses 'Si' in the while condition, but this variable is not defined in the algorithm's notation; it should likely be S'_I or Senv.","section":"Algorithm 1"},{"comment":"The text refers to 'DSAI' in two places where the baseline name 'DIAS' is meant; these typos should be corrected.","section":"IV-A"},{"comment":"The comment 'Initilization' in Algorithm 1 is misspelled and should be 'Initialization.'","section":"Algorithm 1"},{"comment":"The description of S as an element of R^4 with dx, dy, dz and a fourth dimension is confusing; the notation in Eq. (1) suggests a voxel grid whose cells are marked observed or unobserved, so the scene representation should be clarified accordingly.","section":"III-A"},{"comment":"The objective function for OR-MCTS defines a as the action sequence but does not explicitly state that a must respect pick-and-place feasibility within the current observed scene; this constraint should be stated formally.","section":"III-B-2"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work (VPFormer [5] and SS-MCTS [31]) for core components and as baselines, and it omits the directly relevant earlier method [20] from the same group. This is a visibility concern rather than evidence of misconduct, especially because external baselines TSAD and WTL are included. The planning-level success metric and the limited real-robot evidence are the main threats to the paper's central claim; these are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth taking seriously. The new thing is not active sensing or MCTS rearrangement alone, but the tight feedback loop between them: the planner asks the sensor to look at the swept volumes and at relocation regions, and the planner's failures trigger new observations. The OR-MCTS extension that lets objects be placed inside the target swept volume when it helps clear a dependency is a genuine twist on SS-MCTS. The ablations back up the central claim — IAS+FAS beats IAS+SAS, and MAS beats both — so the feedback loop is doing real work. I believe that result.\n\nThe soft spots are where I'd push. First, the success metric in Tables I and II is \"generate a feasible retrieval plan before the scene is entirely observed.\" That is planning success, not execution. The planner's collision checks, swept-volume computations, and grasp selection operate on a geometric model; there is no physics simulation and no control layer. The one detailed real-world demo is nice, but it is one scene among five, and the paper admits the segmentation fails in low light. So the 95% number does not establish reliable physical retrieval. It establishes that the planner can find plans under an optimistic model.\n\nSecond, the paper does not compare against [20], which is the same group's recent neural rearrangement planning for retrieval with an in-hand RGB-D sensor — arguably the closest prior system to this one. It is cited in the related work but omitted from the experiments. For a paper claiming a significant margin over \"state-of-the-art methods\", that omission is hard to justify.\n\nThird, no significance tests. The differences between 95% and 78% look meaningful, but there is no sense of variance across random seeds or scene distributions. For a 100-scenario evaluation, that could be added without pain.\n\nFourth, no code or data. The specific heuristics in the region-specific sensing policy matter, and the paper would be far more reproducible with a release.\n\nNone of this is fatal. The core idea is sound, the ablations are the right kind of evidence, and the authors are transparent about the real-world segmentation limits. The paper deserves peer review, not desk rejection. I would recommend a conditional accept: either report execution-based success (simulated physics or more real-robot trials) or explicitly reframe the claim as planning-level, and add the comparison with [20]. With those changes, the contribution would be solid.","headline":"A genuinely integrated sensing-planning loop with solid ablations, but the headline 95% success rate is a planning metric, not physical retrieval, and the closest prior baseline is missing.","tokens_in":11812,"tokens_out":2728,"would_cite":true,"duration_ms":26287,"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":"Retrieving objects from unknown, cluttered, confined spaces is best treated as one loop in which the planner tells the camera what to look at and the camera tells the planner where it can move things.","keywords":["object retrieval","active sensing","rearrangement planning","Monte-Carlo tree search","confined cluttered environments","in-hand camera","unknown environments","robot manipulation"],"falsifier":"Run the plans returned by this system end-to-end on a physical arm (or in a physics simulator with contact) across the same 100 random confined, cluttered scenes and count actual retrievals; if successful execution falls well below 95% because planned paths collide, grasps fail, or unmodeled object interactions occur, the central claim fails. A cheaper check is to count how often the segmentation step misses an object in low-light clutter, since the paper already reports such failures in real scenes.","tokens_in":10869,"feed_emoji":"🤖","tokens_out":5215,"duration_ms":47939,"temperature":0.7,"pith_summary":"This paper claims that the right way to retrieve a target object from an unknown, cluttered, confined space is to run active sensing and rearrangement planning as one feedback loop, not as separate phases. A robot arm with a single in-hand camera starts sensing, detects the target, and then lets its rearrangement planner decide which still-unobserved regions are worth looking at, observing the regions that would unblock the planner's most promising partial plan. Across 100 randomly generated simulated cabinet-like scenes, the integrated system finds a feasible retrieval plan in 95% of trials, while the best baseline that separates sensing from planning succeeds in 27% to 79% of scenarios. If the result carries to real hardware, it would let service robots fetch objects from refrigerators, cabinets, and shelves without external cameras or exhaustive pre-scanning of the space.","feed_headline":"Planner-guided sensing retrieves objects from clutter 95% of the time","feed_subtitle":"Bidirectional feedback between camera and rearrangement planner beats dense scans in 100 simulated trials and real cabinets.","key_machinery":"The carrying mechanism is the OR-MCTS planner and the three-stage Multi-objective Active Sensing loop. OR-MCTS is a Monte-Carlo-tree-search planner whose nodes are scene configurations linked by pick-and-place actions; it minimizes total relocation distance, treats unobserved voxels as obstacles, and, unlike its predecessor, permits temporary placement of objects inside the target-retrieval swept volume. When a search fails, the region-specific sensing policy clusters the unobserved space, scores each cluster by how much additional feasible relocation area it would open for the remaining blocking objects, and observes the best cluster with a greedy, centroid-focused viewpoint heuristic. That observed volume is fed back into a new search, repeating until a plan is found or the scene is fully seen.","core_discovery":"The paper's central claim is that bidirectional feedback between a multi-objective active sensor and a Monte-Carlo-tree-search retrieval planner is the decisive factor for object retrieval from unknown confined clutter. The sensor has three stages: it maximizes scene coverage until the target is seen, then senses the robot's swept volumes so the planner does not mistake the robot's own future paths for obstacles, and finally, when the planner fails, it senses the unobserved region whose observation would open the most relocation space for the remaining blocking objects. The planner, for its part, is allowed to place non-target objects inside the target's swept volume when that is the only way to reach other blockers, and it returns the most promising failed node to guide the next sensing step. In simulation this system achieves a 95% success rate over 100 random scenes and outperforms its own ablations and three state-of-the-art retrieval planners; in five real cabinet scenes, the paper reports similar average statistics, with one detailed demonstration, and notes that segmentation can fail in cluttered, low-light conditions.","pith_inferences":["The 'sense what the failed plan needs' principle should transfer to other manipulation-under-uncertainty tasks, such as shelf stocking, surgical tool fetch, or disaster debris clearing, wherever an in-hand sensor and a planner share a workspace.","The method's success metric is plan feasibility, not executed retrieval; a natural test is to run the output plans end-to-end in a physics simulator, where swept-volume approximations and grasp feasibility can be checked before committing to hardware.","Performance beyond the tested regime is untested: scenes with more than eight blocking objects, heavy occlusion between objects, or non-convex obstacles could stress the MCTS branching factor and the region-growing cluster scoring, so scaling behavior is the clearest open question.","A direct comparison against an oracle with full scene knowledge would quantify how much of the remaining 5% failure is due to sensing choices rather than rearrangement difficulty."],"forward_implications":["Success no longer requires dense pre-scanning: the 95% success rate is reached with an average of 3.71 viewpoints, whereas dense sensing to 95% coverage succeeds in only 27% of scenes.","Planner feedback is the main driver: removing feedback-based sensing drops success from 95% to 61%, while keeping feedback but removing swept-volume sensing keeps success at 92% but increases failed planning attempts from 1.41 to 3.97.","Allowing temporary placement of objects inside the target's swept volume matters: the same perception pipeline with a planner that forbids this succeeds in 78% rather than 95% of scenes.","The integrated approach finds plans faster and with fewer retries than the baselines, with mean planning time 10.20 seconds versus 14.54 to 18.58 seconds for the other planners."],"supporting_citations":[{"why":"Supplies the imitation-learning viewpoint policy used as the initial sensing stage that detects the target object.","marker":"[5]"},{"why":"Provides the Monte-Carlo-tree-search rearrangement backbone that OR-MCTS extends with target-swept-volume placement and sensing feedback.","marker":"[31]"},{"why":"Generates the collision-free robot paths whose swept volumes identify the non-target objects blocking target retrieval.","marker":"[33]"},{"why":"Produces candidate six-degree-of-freedom grasps on the target object, from which the planner selects the lowest-swept-volume path.","marker":"[32]"},{"why":"Clusters unobserved regions during feedback-based sensing so the planner can score each cluster by the relocation space it would open.","marker":"[36]"},{"why":"Serves as a retrieval-planning baseline that relocates one object per MCTS stage with a random relocation-region policy.","marker":"[17]"},{"why":"Serves as a retrieval-planning baseline that divides space into valid and invalid relocation regions and minimizes newly invalid regions.","marker":"[18]"}],"fun_headline_variants":["Bidirectional sensing and planning retrieves objects 95% of the time","Sensor and MCTS planner trade feedback to beat clutter","95% retrieval: sensing and rearrangement planner co-adapt","Integrated active sensing and MCTS planning retrieve 95%","Bidirectional feedback loop clears clutter for object grasp"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a plan that passes the planner's swept-volume, motion, and grasp checks will also succeed when physically executed; the simulated success rate measures plan feasibility before the scene is fully observed, and real-world support is one detailed demonstration among five, with segmentation failures acknowledged in cluttered, low-light scenes.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional sensing and planning retrieves objects 95% of the time","Sensor and MCTS planner trade feedback to beat clutter","95% retrieval: sensing and rearrangement planner co-adapt","Integrated active sensing and MCTS planning retrieve 95%","Bidirectional feedback loop clears clutter for object grasp"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2020,"prompt_tokens":949,"completion_tokens":1071,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":990}},"tokens_in":565,"tokens_out":1071,"duration_ms":8485,"temperature":1.0,"reasoning_tokens":990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:12:07.360025+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the plans returned by this system end-to-end on a physical arm (or in a physics simulator with contact) across the same 100 random confined, cluttered scenes and count actual retrievals; if successful execution falls well below 95% because planned paths collide, grasps fail, or unmodeled object interactions occur, the central claim fails. A cheaper check is to count how often the segmentation step misses an object in low-light clutter, since the paper already reports such failures in real scenes.","supporting_citations":[{"cited_title":"Robot active neural sensing and planning in unknown cluttered environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the imitation-learning viewpoint policy used as the initial sensing stage that detects the target object."},{"cited_title":"Multi-Stage Monte Carlo Tree Search for Non-Monotone Object Rearrangement Planning in Narrow Confined Environments","cited_arxiv_id":"2305.17175","evidence_quote":"Provides the Monte-Carlo-tree-search rearrangement backbone that OR-MCTS extends with target-swept-volume placement and sensing feedback."},{"cited_title":"Contact- graspnet: Efficient 6-dof grasp generation in cluttered scenes,","cited_arxiv_id":null,"evidence_quote":"Produces candidate six-degree-of-freedom grasps on the target object, from which the planner selects the lowest-swept-volume path."},{"cited_title":"Image segmentation using seeded region growing,","cited_arxiv_id":null,"evidence_quote":"Clusters unobserved regions during feedback-based sensing so the planner can score each cluster by the relocation space it would open."},{"cited_title":"Object rearrangement planning for target retrieval in a confined space with lateral view,","cited_arxiv_id":null,"evidence_quote":"Serves as a retrieval-planning baseline that relocates one object per MCTS stage with a random relocation-region policy."},{"cited_title":"Where to relocate?: Object rearrangement inside cluttered and confined environments for robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"Serves as a retrieval-planning baseline that divides space into valid and invalid relocation regions and minimizes newly invalid regions."}],"review_version":1}