{"id":"4d6671e9-00a6-46e1-9a24-66433e0cdc47","arxiv_id":"2508.12916","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A robot using a single wrist-mounted RGB-D camera can retrieve objects in cluttered scenes by combining a dynamic scene graph, VLM-driven viewpoint selection, interactive exploration, and grasping, with reported success rates of 70-90%.","lead":"RoboRetriever is a robot system that finds and retrieves objects using only one wrist-mounted camera and natural language instructions. It builds a dynamic scene memory, decides where to look, and interacts with objects to uncover hidden targets, reporting success on real tabletop tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline success rates are computed without disclosed denominators: each task category has three variations but the paper never states how many rollouts each variation received, so 70–90% may reflect only a handful of successes and cannot support the claimed cross-task generalization.","rationale":"Good-faith reading: the contribution is an integrated system whose central claim is capability without task-specific priors. For that claim to hold, the evidence must show reliable performance across diverse tasks. The architecture is plausible, and the qualitative rollout shows real coordination of active perception, interactive perception, and manipulation. However, the strongest evidence is Table 2, and the paper omits the one number needed to interpret it: the number of rollouts. 'Each task category has three variations' might mean three total episodes per category or three conditions each replicated many times; the text never says. With no error bars and no raw logs, the success percentages are not auditable. This is the weakest link because the novel prompting scheme, scene graph, and action integration all serve these outcomes; if the outcomes rest on a few selected rollouts, the central claim collapses to a demonstration, not a general result. The reader's VLM concern is valid but secondary: closed-model reliance affects reproducibility and robustness, but even a fully open model would not fix an underpowered or undisclosed evaluation. Conversely, if the trial counts are large and raw logs are released, the closed-VLM issue becomes a reproducibility caveat rather than a refutation. I therefore treat the evaluation denominator as the load-bearing concern. The proposed check is simple, requires no additional hardware runs, and directly settles whether Table 2 supports the claimed generalization. If n is small, the verdict should remain conditional on additional trials across more environments; if n is adequate and logs agree, the central claim is far better supported. Hence the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":11483,"tokens_out":4143,"duration_ms":46143,"concrete_test":"Obtain from the authors the raw per-variation run logs and the number of rollouts n per task category, then recompute exact binomial 95% confidence intervals for each success rate in Table 2. If n is 3 or smaller, the headline 90% interval includes 50% or below and Table 2 cannot establish the claimed cross-task generalization. If n is 30 or more per category and the per-variation logs confirm the aggregate rates, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 is the sole quantitative evidence for the central claim that RoboRetriever 'adapts across diverse tasks and environments without task-specific priors.' The setup section says each category has three variations in instruction, object number, type and layout, but it never reports the number of rollouts per variation, total trials per category, or per-variation outcomes. If one run was executed per variation, n=3 per category: a reported 90% success rate is 2.7/3, with a binomial 95% confidence interval roughly 9–99% for 2/3 or 29–100% for 3/3, and the 70% Compositional result is about 2/3. With such small samples, the large margin over baselines (0–40%) is not statistically meaningful. The generalization claim is further strained because all trials share one tabletop, one arm, one camera, and hand-authored variations. This is an evidentiary gap rather than an internal contradiction, but it is load-bearing: the paper's headline numbers, not the architecture alone, carry the generalization claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RoboRetriever, a framework for object retrieval with a single wrist-mounted RGB-D camera and free-form natural-language instructions. The system builds and incrementally updates a dynamic hierarchical scene graph through a grounding module (GPT-4o, DINO-X, SAM), a memory module that includes hypothetical Unknown nodes, a supervisor module based on GPT-o3, and an action module that coordinates active perception, interactive perception, and manipulation. Active perception is driven by a novel visual prompting scheme in which GPT-o3 selects among sampled camera directions and poses. The method is evaluated on six real-world tabletop task categories against RoboEXP, AP-VLM, and a GPT-o3 baseline, reporting success rates of 70--90% and object discovery rates up to 95%, along with ablations and a supplementary video of human-intervention scenarios.","tokens_in":11709,"tokens_out":3859,"duration_ms":41838,"significance":"If the reported results hold, RoboRetriever would be a meaningful advance for single-camera, language-driven object retrieval in partially observable scenes, and the integration of active and interactive perception within one modular framework is genuinely interesting. The dynamic scene graph with Unknown nodes and the task-aware visual prompting scheme for 6-DoF camera control are plausible contributions. However, the current evidence is insufficient to support the strong generalization and robustness claims: the central quantitative results lack trial counts and uncertainty measures, the closest language-grounded baselines are absent, and the heavy reliance on undisclosed closed-model prompts makes independent verification impossible. The paper is promising but needs substantial empirical and transparency revisions.","major_comments":[{"comment":"The central generalization claim rests on Table 2, but the manuscript never states the number of rollouts per variation or per task category. The setup says each category has three variations, so if only one rollout was executed per variation, the reported 90% success rate could correspond to 3/3 successes with a binomial 95% confidence interval of roughly 29--100%, and the 70% Compositional Reasoning result could be 2/3 with an even wider interval. Please report the exact number of trials per variation and per category, the per-variation outcomes, and confidence intervals, and if the trial counts are this small, temper the headline success-rate claims accordingly.","section":"Experiments (Table 2)"},{"comment":"The evaluation omits the closest language-grounded interactive search baselines, such as the language-grounded dynamic scene graph approach of Honerkamp et al. and CuriousBot, both of which are cited in the paper. In addition, the GPT-o3 baseline is asymmetric: it receives ground-truth interactive perception and manipulation actions and is not given the memory module, which makes the large performance margin less informative. Please add at least one recent language-grounded interactive search baseline, or explicitly justify why those methods are not comparable, and equalize the information available to baselines where possible.","section":"Experiment Setup (Baselines)"},{"comment":"The entire pipeline depends on closed commercial models (GPT-4o and GPT-o3) for instance matching, semantic attribute updates, relation inference, Unknown-node placement, supervisor decisions, and camera pose selection, but the paper provides no exact prompts, API versions, sampling parameters, or temperature settings, and it does not audit VLM failures. Because no code or data are released, these omissions prevent independent reproduction and make the results vulnerable to model-version drift. Please include the full prompts and model configuration in an appendix, and add a failure analysis with representative examples of VLM errors and their effect on task outcomes.","section":"Action module / Grounding module"},{"comment":"The claim that the method adapts 'without task-specific priors' is strained by several unreported tunable parameters in the active perception module: the sphere radius and distance factor, the number of sampled candidate directions N, and the number of sampled candidate poses M. None of these values is given, and no sensitivity analysis is provided. Please report these parameters and show that the results do not depend critically on their choice.","section":"Action module (Active perception)"},{"comment":"The claimed robustness under human interventions is supported only by a statement that 'more details can be found in the supplementary video,' with no quantitative success rate, trial count, or failure description in the paper. Similarly, the ablation results in Figure 6D and 6E are presented without error bars or explicit numbers of rollouts. Please provide quantitative results with trial counts for the ablation conditions and for human-intervention trials.","section":"Ablation Study (Human intervention)"}],"minor_comments":[{"comment":"There is a typo in the phrase 'generative rather than discrimitive approach'; it should be 'discriminative'.","section":"Comparison with baselines"},{"comment":"The table uses '/' for GED entries of AP-VLM and GPT-o3; please explain whether GED was not computed for these baselines and clarify the exact computation of the ground-truth and predicted scene graphs.","section":"Table 2"},{"comment":"The visual legend for the scene graph is difficult to follow because omitted nodes are indicated with ellipses and several node types overlap; please enlarge the figure and provide a clearer key.","section":"Figure 5"},{"comment":"The paper should state explicitly that all trials share the same tabletop, robot arm, camera, and gripper, and discuss the resulting limitations for generalization to other embodiments and environments.","section":"Experiment Setup"},{"comment":"The Object Discovery Rate is defined as 'the percentage of discovered objects out of all objects in the environment,' but the denominator is ambiguous for objects that are never visible from any viewpoint; please specify how the ground-truth object inventory was obtained.","section":"Metrics"}],"recommendation":"major_revision","confidential_remarks":"The novelty is plausible and the system is ambitious, but the empirical evidence is not yet at the level needed for the claims made. The main issues are missing trial counts, missing prompts and API versions, absent stronger baselines, and no code or data release. I would like to see the trial counts and prompts before judging whether the reported margins are real; no misconduct is suspected, but the current text overstates the strength of the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you work on VLM-driven robot manipulation. The system is real: one wrist-mounted RGB-D camera, a dynamic hierarchical scene graph that includes \"unknown\" nodes for unexplored regions, and a two-stage visual prompting scheme that gets GPT-o3 to pick 6-DoF camera poses. The integration is new, the unknown-node idea is neat, and the paper is honest about its building blocks. It also shows the system on a physical arm across six task categories, including human intervention and sequential instructions, which is more than many papers in this area do.\n\nThe central claim is that the method generalizes across tasks and environments without task-specific priors. That claim is under-supported. The evaluation section never reports how many rollouts each of the three variations per category received, so Table 2's 70–90% success rates could be 2/3 or 9/10. With n = 3, the confidence intervals span nearly the whole range, and the margin over the baselines is not statistically meaningful. The setup is also a single tabletop, one arm, and hand-authored variations, so the \"diverse environments\" part of the claim is mostly aspirational. The GPT-o3 baseline is given ground-truth interaction actions, which makes its failure notable, but it is still a weak baseline; the closest language-grounded search systems (e.g., Honerkamp et al. 2024) are cited but not compared.\n\nThe other soft spot is reproducibility. No code, no data, no exact prompts, no API versions. The whole pipeline leans on GPT-4o and GPT-o3 for matching, attribute updates, relation inference, and pose selection, so the paper as written cannot be re-run or audited. That is a real limitation, not a minor one.\n\nWhat the paper does well is the architecture and the qualitative demonstrations. The ablation figures suggest active perception and memory both matter, and the system's ability to recover from occlusion by opening drawers and moving objects is impressive. But the quantitative evidence is too thin to justify the strong generalization claims.\n\nFor peer review: yes, it deserves a serious referee, because the system is nontrivial and the problem is important. But the referee should demand trial counts, per-variation results, error bars, and an artifact release before acceptance. I would not cite the numbers as published, though I might cite the prompting idea after it is properly evaluated.\n\nBring it to reading group, but pair it with a discussion of what counts as adequate evidence for a systems claim in embodied AI.","headline":"A real single-camera retrieval system with a clever prompting scheme, but the headline numbers rest on undisclosed trial counts and the closed-model dependency makes the evaluation hard to trust as reported.","tokens_in":12224,"tokens_out":2336,"would_cite":false,"duration_ms":26637,"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":"RoboRetriever claims that a robot with only one wrist-mounted RGB-D camera can retrieve hidden, occluded, or semantically specified objects by actively choosing viewpoints and physically interacting with the scene, reporting 70–90%…","keywords":["single-camera object retrieval","active perception","interactive perception","dynamic scene graph","vision-language models","visual prompting","robot manipulation","partial observability"],"falsifier":"A concrete test: replicate the Hidden Inside and Compositional Reasoning tasks while logging every GPT-o3 decision, and count (a) camera poses that collide or point away from the scene, (b) object merges that conflate distinct instances, and (c) rollouts that need human approval or correction; if any of these occur in more than a small fraction of trials, the claimed single-camera autonomy is not established.","tokens_in":11282,"feed_emoji":"🤖","tokens_out":6125,"duration_ms":60118,"temperature":0.7,"pith_summary":"The paper claims that a robot equipped with only a single wrist-mounted RGB-D camera can carry out free-form object-retrieval instructions in cluttered, partially observable scenes by combining active perception (deliberately moving the camera) with interactive perception (opening, closing, moving, or rotating objects). To do this it maintains a dynamic hierarchical scene graph that accumulates object semantics, geometry, and spatial relations over time, and it uses a reasoning vision-language model to decide, from rendered views of the scene point cloud, where to look next and what to do next. On six real-world task categories, including hidden objects, recursive search, semantic targeting, and sequential instructions, the paper reports success rates of 70–90% and object discovery rates up to 95%, with large gaps over fixed-camera, grid-scan, and pure-VLM baselines. The significance would be a step toward robotic retrieval that needs no fixed or multi-camera setup, no task-specific camera heuristics, and no retraining for a new instruction.","feed_headline":"One wrist camera finds hidden objects up to 90 percent of the time","feed_subtitle":"The robot builds a scene graph, picks its own viewpoints, opens drawers, and clears clutter to follow a natural-language request.","key_machinery":"The central object is the dynamic hierarchical scene graph, defined as $G=(V,E)$, where each node carries a cropped image history, an accumulated point cloud, semantic attributes (fine-grained name, movable/static flag, confidence, occluded flag, partial-view flag, and a free-form description), and edges encode five relations: behind, belong, inside, on, under. Relations pointing into unexplored regions trigger 'Unknown' nodes that hypothesize objects behind occluders or inside containers, which is what lets the supervisor formulate exploration goals. The action-driving mechanism is the visual prompting scheme: the system samples candidate camera directions on a virtual sphere centered on the target object, renders canonical front/left/right views of the point cloud, and has GPT-o3 choose the best direction and then the best pose, keeping the camera aimed at the object center; this turns pose selection into a grounded multiple-choice decision rather than free-form coordinate generation.","core_discovery":"The central claim is that active and interactive perception can be unified under a single-camera constraint: instead of assuming a fixed or multi-camera setup with full scene visibility, the robot builds a dynamic hierarchical scene graph from its wrist-camera observations, continuously updates that graph as it moves and interacts, and lets a reasoning vision-language model decide both which object to pursue and which action to take next. The paper's reported results show that this integration succeeds on tasks where each ingredient alone fails: active perception alone cannot open a closed drawer, interactive perception alone cannot choose where to look, and a VLM without grounded scene memory hallucinates camera poses. The framework is presented as class-agnostic and task-agnostic, adapting to new instructions and environments without task-specific priors.","pith_inferences":["Inference: because pose selection is a discrimination task over sampled candidates rather than generative coordinate regression, the scheme may work with any VLM strong enough to compare rendered views; swapping GPT-o3 for an open-weight model would test this directly.","Inference: the 'Unknown' nodes function as explicit exploration frontiers, so graph-edit distance may serve as a measure of information gain; a next-action policy could optimize expected graph change instead of relying on the VLM's free choice.","Inference: the supervisor's decision loop is not tied to tabletop arms; the same scene-graph-plus-prompt pattern could drive mobile manipulators or dual-arm systems, where each camera move is still a 6-DoF viewpoint selection.","Inference: the reported numbers average over successes; a stricter evaluation would separate perception accuracy from planning accuracy by replaying logged VLM decisions against ground-truth scene states, which the paper does not report."],"forward_implications":["Retrieval no longer requires full scene visibility: a single moving camera plus physical interaction can expose objects that no fixed view captures.","New instructions do not require changing perception code: the same supervisor prompt consumes the current scene graph and produces a plan for any object named in natural language.","Memory of past actions and objects pays off in sequential tasks: later instructions can reuse the accumulated graph instead of re-exploring the scene.","The reported gap over the GPT-o3 baseline suggests that grounding viewpoint choice in rendered point-cloud views substantially reduces hallucinated camera poses.","The combination of active and interactive perception handles occlusion modes that either alone cannot: the robot can look behind, under, and inside, and can move obstructions when vision alone is insufficient."],"supporting_citations":[{"why":"Supplies RoboEXP, the interactive-exploration scene-graph baseline that the paper compares against and extends with active perception.","marker":"(Jiang et al. 2024)"},{"why":"Supplies AP-VLM, the grid-based active-perception baseline that uses a wrist camera but cannot interact with the scene.","marker":"(Sripada et al. 2024)"},{"why":"Supplies DINO-X, the open-world detector used in the grounding module to obtain object bounding boxes.","marker":"(Ren et al. 2024)"},{"why":"Supplies Segment Anything (SAM), used to produce segmentation masks that extract object point clouds from RGB-D images.","marker":"(Kirillov et al. 2023)"},{"why":"Supplies AnyGrasp, the grasp-perception method used to execute the final manipulation of the retrieved object.","marker":"(Fang et al. 2023)"},{"why":"Supplies ZeroMatch, the point-cloud registration algorithm that merges new observations into each object's accumulated point cloud.","marker":"(Jiang et al. 2025)"}],"fun_headline_variants":["One wrist camera, 90% hidden-object retrieval","RoboRetriever: single-camera active perception for hidden finds","Robot uses scene graph and one camera to retrieve objects","Active and interactive perception unified under one camera","Wrist-camera robot retrieves objects without multi-camera setups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline trusts closed-source vision-language models GPT-o3 and GPT-4o to make every semantic decision—matching objects across views, updating attributes, inferring relations, adding unknown nodes, choosing what to do next, and picking the next camera pose—so if those models hallucinate, drift, or become unavailable, the robot's autonomy collapses.","fun_headline_variants_meta":{"raw":{"variants":["One wrist camera, 90% hidden-object retrieval","RoboRetriever: single-camera active perception for hidden finds","Robot uses scene graph and one camera to retrieve objects","Active and interactive perception unified under one camera","Wrist-camera robot retrieves objects without multi-camera setups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":1998,"prompt_tokens":943,"completion_tokens":1055,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":974}},"tokens_in":559,"tokens_out":1055,"duration_ms":10993,"temperature":1.0,"reasoning_tokens":974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:17:24.112568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: replicate the Hidden Inside and Compositional Reasoning tasks while logging every GPT-o3 decision, and count (a) camera poses that collide or point away from the scene, (b) object merges that conflate distinct instances, and (c) rollouts that need human approval or correction; if any of these occur in more than a small fraction of trials, the claimed single-camera autonomy is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies RoboEXP, the interactive-exploration scene-graph baseline that the paper compares against and extends with active perception."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies AnyGrasp, the grasp-perception method used to execute the final manipulation of the retrieved object."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies ZeroMatch, the point-cloud registration algorithm that merges new observations into each object's accumulated point cloud."}],"review_version":2}