{"id":"5ac1dcaa-a525-4247-92db-b1ddadf45ba5","arxiv_id":"2504.15278","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single video of a static indoor scene can be turned into a photorealistic, interactive virtual environment with working articulated objects, usable for games and robot learning.","lead":"DRAWER automatically converts a video of a room into a 3D virtual copy with openable drawers and cabinets that can be imported into game engines and robot simulators. It aims to replace manual 3D modeling and reduce the cost of training robots in realistic virtual spaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is that joint types and hinge axes can be inferred from static closed furniture; Table 4's malformed Ours row leaves its accuracy unverifiable, and no robustness analysis covers unusual mechanisms.","rationale":"The reader identified the same weak point: articulation from static observations. My reading agrees, and adds that the paper's own Table 4, if taken at face value, may show only 58/78 correct articulation types for DRAWER, which would directly undercut 'significantly outperforms prior art' and 'accurate articulation'. Even allowing that the table is malformed in transcription, the absence of per-object results, error bars, and released code means the claim cannot be independently checked. The dual-scene representation and straight-through estimator are plausible and the qualitative demos are suggestive, but they do not establish the central assertion about articulation accuracy and sim-to-real utility. The robot transfer experiment is qualitative only, with no measured success rates, so it does not rescue the claim. I would keep the reader's CONDITIONAL verdict: accept only if code/data, corrected tables, and a diverse-mechanism articulation evaluation are provided. No ad hominem is intended; this is an evidence-and-reproducibility concern.","tokens_in":17005,"tokens_out":4991,"duration_ms":48136,"concrete_test":"Run DRAWER's articulation module on a held-out set of closed articulated objects spanning diverse mechanisms (front-hinged cabinet, sliding door, double door, push-to-open, hidden hinge), with ground-truth joint types and axes measured by physically opening each object. Report per-mechanism precision/recall for type prediction and angular error for predicted hinge axes. Re-run Table 4 with corrected column alignment on the same objects; if Ours correctly predicts fewer than ~95% of types, or if accuracy on non-front-hinged mechanisms drops well below the standard-drawer category, the central articulation claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central interactivity claim depends on Sec. 3.3's physical reasoning: 3DOI and GPT4o infer each object's articulation type and hinge axis from a static video of closed furniture, with no observed motion. Many real mechanisms (sliding doors, double doors, push-to-open, hidden hinges) are underdetermined from static appearance alone, so a correct result is not guaranteed by the perception stack. The evidence is too thin to carry this weight: Table 4 reports only aggregate counts on 78 objects in six kitchens, with no per-scene or per-mechanism breakdown, no failure analysis, and no error bars. Worse, as printed the table is internally inconsistent: the 3DOI row is '80 78 59 57 0.861' while the Ours row is '78 58 0.994', with one column missing; under the most natural alignment Ours correctly identifies only 58/78 articulation types (74%). If that reading is right, the paper's own table contradicts 'accurate articulation' rather than supporting it; if it is wrong, the table cannot be checked because code and data are unreleased. The subsequent game and real-to-sim-to-real demos all presuppose correct joints, so this uncertainty is load-bearing for the paper's central claim, not a peripheral evaluation issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DRAWER, a framework that takes a video of a static indoor scene and produces an interactive digital twin. It combines a neural SDF for geometry with Gaussian splatting anchored to the SDF's zero level set for appearance, and then uses foundation models (3DOI, GPT4o) to infer articulation types and hinge axes, fits category-level shape templates to complete hidden regions, and composes the articulated objects back into the scene. The authors evaluate rendering quality, articulation inference, and motion simulation on six kitchen scenes, and demonstrate an Unreal Engine game and a real-to-sim-to-real robot drawer-opening task.","tokens_in":17313,"tokens_out":5364,"duration_ms":43508,"significance":"If validated, DRAWER would be a useful step toward automatic creation of photorealistic, physically interactive digital twins from casual video, with clear applications in gaming and robot learning. The dual representation with face-anchored Gaussians and a straight-through estimator is a sensible way to couple appearance with geometry, and the paper attempts to evaluate articulation inference and motion simulation rather than only novel-view synthesis. The main weaknesses are the small evaluation set, internally inconsistent result tables, and the unverified reliance on static-appearance articulation inference, all of which currently prevent the central claims from being accepted as demonstrated. The paper also ships no code, data, or supplementary material, limiting reproducibility.","major_comments":[{"comment":"Table 4 is internally inconsistent: the 3DOI row has five entries (80, 78, 59, 57, 0.861) while the Ours row has three (78, 58, 0.994), so it is impossible to determine the correct articulation count for DRAWER. Under the most natural column alignment, Ours correctly identifies only 58 of 78 articulation types (74%), which would contradict the claim of accurate articulation; under an alternative alignment the table cannot be checked because code and data are not released. Please provide a corrected table with the full column set, per-scene and per-mechanism breakdowns, and error bars or confidence intervals for the 78-object evaluation.","section":"Table 4, Sec. 4.2"},{"comment":"Table 3 is also malformed: the Digital Cousin row contains only four numeric entries (160, 75, 46.9, 60.0) under a five-column header, and the reported precision/recall values are inconsistent with an integer correct count. Moreover, the asterisk notes that URDFormer and Digital Cousin were evaluated by taking the best result over multiple runs per scene, while no such handicap is reported for DRAWER; this makes the comparison unfair. Please supply the corrected table and state whether DRAWER also used multiple predictions.","section":"Table 3, Sec. 4.2"},{"comment":"The Introduction claims DRAWER 'significantly outperforms prior art across all metrics,' but Table 1 shows 3DGS achieves higher PSNR (30.42 vs 27.80), SSIM (0.954 vs 0.912), and LPIPS (0.126 vs 0.159). The claim is therefore contradicted by the paper's own results unless 'metrics' is restricted to interactive compatibility. Please qualify the claim and discuss the trade-off explicitly.","section":"Sec. 1, Table 1"},{"comment":"The articulation module infers articulation types and hinge axes from static observations of closed furniture using 3DOI, GPT4o, and a VLM arbiter. Many real mechanisms (sliding doors, double doors, push-to-open, hidden hinges) are underdetermined from static appearance, and the manuscript reports no per-mechanism accuracy, no failure analysis, and no robustness test on unusual furniture. Since the downstream game and real-to-sim-to-real demos presuppose correct joints, this omission is load-bearing for the central contribution. Please add a confusion matrix or per-mechanism breakdown and a discussion of failure modes.","section":"Sec. 3.3, Table 4"},{"comment":"The empirical evaluation rests on six manually captured kitchen scenes with no error bars or variance estimates, and the code, data, and trained models are not released. Single-run results on six scenes cannot support the broad claims of generalization in Sections 1 and 4. Please add run-to-run variance (e.g., across random seeds or scene subsamples) and, at minimum, release the evaluation data and annotations.","section":"Sec. 4.1, Sec. 4.2"}],"minor_comments":[{"comment":"The title appears as 'DRA WER' in the running header and in the abstract title; please correct the spacing throughout.","section":"Title and running header"},{"comment":"The asterisks on BakedSDF* and 2DGS* are not explained in the caption or text; please define them.","section":"Table 1 caption"},{"comment":"The caption states that for URDFormer and Digital Cousin 'we select the best results from multiple image runs,' but no equivalent selection is documented for DRAWER; please clarify whether DRAWER also used multiple runs.","section":"Figure 6 caption"},{"comment":"The reported EMD values (1.41e-5 vs 17.7e-5) lack units and normalization details; please specify the trajectory length, scale, and the number of points used in the Earth Mover's Distance computation.","section":"Sec. 4.2, EMD comparison"},{"comment":"The paper repeatedly refers to the supplementary material for details of the SDF losses, Gaussian parameterization, rigid-object pipeline, and additional experiments, but no supplementary document is included in the arXiv submission; please include it.","section":"General"},{"comment":"The real-to-sim-to-real experiment reports no quantitative success rates or comparison to a manual-articulation baseline, so the claim that DRAWER 'eliminates the need for tedious manual effort' is supported only by qualitative videos; please provide quantitative metrics.","section":"Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistencies in Tables 3 and 4 are concerning and should be addressed before further review; they suggest the experimental reporting was rushed. The baseline handicap in Table 3 and the absence of error bars further weaken the empirical case. Please also verify the claim in Section 1 against Table 1. I do not see evidence of misconduct, but the paper needs substantial revision to make the results verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nYou should know two things about DRAWER. It's a serious systems paper with a genuinely new dual representation, and its articulation evaluation is too thin to support the headline claims as printed. The face-anchored Gaussians on an SDF mesh with a straight-through estimator is a sensible design, and the static-video articulation pipeline is genuinely new at scene level. The demos—an Unreal Engine game and a real-to-sim-to-real robot drawer-opening run—show real integration. The ablation in Table 2 is clean and each component earns its place. The idea of using a VLM to arbitrate between 3DOI and GPT4o is pragmatic, and the amodal shape estimation with compositional templates addresses a real problem.\n\nThe soft spots are in the evaluation. Table 4, which is the only quantitative support for articulation inference, is malformed. The 3DOI row has five numbers and the Ours row has three. If read left-to-right, Ours says 58 correct out of 78 total (74%), which contradicts 'significantly outperforms' 3DOI's 78/80. If that reading is wrong, the table cannot be checked because no code or data are released. This is load-bearing, not cosmetic: the entire interactivity claim rests on getting the joints right, and we are given no per-scene breakdown, no error bars, and no failure analysis for unusual mechanisms like sliding doors or push-to-open. The baseline comparison is also uneven: URDFormer and Digital Cousin get best-of-multiple-runs, DRAWER gets one run. The robot transfer section is qualitative only, with no measured success rates. Six kitchens with 78 objects is a small evaluation set.\n\nNone of this sinks the paper; the reconstruction results and the end-to-end system are real. But as printed, the quantitative support for the central articulation claim is not there.\n\nThis paper is for anyone working on interactive scene reconstruction, content creation, or sim-to-real transfer. It deserves a serious referee, not a desk reject. My recommendation: send it to review, and condition acceptance on release of code and data, a corrected Table 4 with per-scene counts and error bars, and baseline comparisons on equal footing. If the corrected articulation numbers are as low as 74%, the paper becomes a reconstruction and integration contribution, which is still worth publishing, but the claims need to be scaled back.","headline":"A serious systems paper with a genuinely new dual representation and a static-video articulation pipeline, but the articulation evaluation is too thin and Table 4 is internally inconsistent as printed.","tokens_in":17860,"tokens_out":3201,"would_cite":false,"duration_ms":28513,"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":"DRAWER claims that a single static video of an indoor scene can be turned automatically into a photorealistic, interactive digital twin whose drawers and doors open with physically correct motion.","keywords":["digital twin","3D reconstruction","Gaussian splatting","neural signed distance field","articulation estimation","interactive simulation","real-to-sim-to-real","novel view synthesis"],"falsifier":"Take a held-out set of furniture with articulation mechanisms that are not visually obvious from a closed state, such as push-to-open drawers, one-way swing cabinets with hidden dampers, or bifold doors, run DRAWER on a static video, and check whether the predicted joint type and axis let the simulated opening follow the real motion within the paper's trajectory-EMD threshold.","tokens_in":16817,"feed_emoji":"🎮","tokens_out":10314,"duration_ms":92520,"temperature":0.7,"pith_summary":"DRAWER aims to show that a single video of a static indoor scene—no interactions captured—is enough to build a photorealistic, interactive digital twin in which drawers and doors can be opened, objects moved, and novel views rendered in real time. Its central claim is that accurate geometry and photorealistic appearance do not have to be traded off: a neural signed distance field supplies the surface, and Gaussian splats anchored to that surface supply the pixels, so the same mesh can serve both rendering and physics. The paper further claims that articulation—joint type and hinge axis for closed cabinets and drawers—can be inferred from appearance with foundation-model perception, then refined with 3D grounding. If these claims hold, content creation for games and robot training could move from hand-modeled assets to automatically generated, physically accurate replicas of real rooms.","feed_headline":"Video of a kitchen becomes a playable, physics-real digital twin","feed_subtitle":"System infers how drawers and doors move from static video alone, then lets you open them in simulation.","key_machinery":"The load-bearing object is the dual scene representation: a neural signed distance field (a function mapping each 3D point to a signed distance from the nearest surface) that is extracted as a mesh to provide collision geometry, plus a set of 3D Gaussian splats spawned at mesh-face centroids, initialized in scale and rotation to the face, and constrained to move only within that face and a limited distance along the normal. A straight-through estimator makes the resulting clipping operation differentiable, so the splats can be trained with the usual photometric loss. The second mechanism is the articulation pipeline: object masks from open-vocabulary segmentation are fused onto the mesh, a foundation model proposes joint types and axes, a vision-language model re-estimates and arbitrates, and differentiable rendering of compositional shape templates aligns the completed objects with the observed partial geometry; hidden interior regions receive plausible material texture from a conditional diffusion model.","core_discovery":"The paper reports that DRAWER converts a single static video into an interactable digital twin by coupling a neural signed distance field with 3D Gaussian splatting in a dual scene representation, and it identifies this coupling as the reason the reconstruction is simultaneously geometry-accurate and photorealistic. Anchoring Gaussians to the SDF's zero level-set, while allowing them limited motion on their parent faces and using a straight-through estimator to keep training differentiable, suppresses floating splats and keeps rendered appearance aligned with collision geometry. On top of that geometry, the articulation module infers joint types and revolute axes from foundation-model predictions and multi-view mask fusion, then fits compositional shape templates to complete hidden interior shapes and textures them with a conditional diffusion model. The paper demonstrates the result by importing the twins into a game engine and by training a drawer-opening policy in simulation that transfers to a real robot, reporting articulation precision of 97.2 percent and a rotation-axis accuracy metric of 0.994 on six captured kitchens.","pith_inferences":["The same static-video-to-interactive-twin recipe should extend to historical interiors, architectural scans, or museum exhibits where physically opening the original is impossible, turning still documentation into walkable, openable virtual spaces.","Because the coupling constraint is generic, the dual representation could be reused for other sensor modalities or for editable scene editing after object removal, where appearance must follow geometry changes.","The reported evaluation covers six kitchens; a direct next experiment is to run the pipeline on furniture with non-standard mechanisms such as push-to-open latches, sliding shelves, or two-door cabinets with hidden center pivots, to see where foundation-model articulation priors break.","A quantitative real-to-sim-to-real study with measured success rates across multiple scenes and object categories would strengthen the robotics claim beyond the reported qualitative demonstration."],"forward_implications":["A static video alone becomes sufficient input for building an editable, physically interactive environment, removing the need to capture scenes before and after every interaction or to hand-specify joints.","Rendering quality stays close to the best splatting baseline (PSNR 27.80, SSIM 0.912, LPIPS 0.159) while the SDF mesh provides a collision-ready geometry, so what the agent sees and what it can touch remain aligned.","Simulated articulated trajectories match real ground-truth motion with an EMD of $1.41 \\times 10^{-5}$, an order of magnitude lower than a video-diffusion baseline's $17.7 \\times 10^{-5}$.","The exported twins plug into standard game-engine and robot-simulator pipelines, so game levels and robot training environments could be produced from casual video capture rather than manual 3D modeling."],"supporting_citations":[{"why":"Supplies the 3D Gaussian splatting representation that DRAWER uses for real-time photorealistic appearance.","marker":"[32]"},{"why":"Provides the neural surface reconstruction method whose extracted mesh supplies accurate geometry and anchors the splats.","marker":"[99]"},{"why":"Surface-aligned Gaussian splatting baseline used in the quantitative rendering comparison to isolate the dual representation's contribution.","marker":"[27]"},{"why":"Supplies per-object articulation and hinge predictions from static images that DRAWER validates and refines with 3D geometry.","marker":"[69]"},{"why":"Prior pipeline that builds articulated simulation environments from images, serving as the main comparison for articulation understanding.","marker":"[10]"},{"why":"Prior automated digital-cousin creation method used as a second baseline for scene-level interactive reconstruction quality.","marker":"[11]"},{"why":"Closest prior system reconstructing an interactable scene from a single video; DRAWER extends it with articulation and the dual representation.","marker":"[89]"},{"why":"Controllable video-generation baseline that receives privileged segmentation and motion prompts yet produces less accurate articulated motion.","marker":"[35]"},{"why":"The diffusion-policy architecture used to train the drawer-opening controller in simulation before real-robot transfer.","marker":"[105]"}],"fun_headline_variants":["DRAWER: From a single video to a playable, photorealistic digital twin","One static video becomes an interactive 3D world with movable parts","DRAWER reconstructs scenes and articulates objects from video alone","Video in, interactive twin out: DRAWER fuses geometry and appearance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central bet is that a closed drawer or cabinet's articulation type and hinge axis can be inferred correctly from static appearance plus common-sense priors, without ever seeing the object move; if that inference is wrong, the simulated joint is wrong and the interactive twin is inaccurate.","fun_headline_variants_meta":{"raw":{"variants":["DRAWER: From a single video to a playable, photorealistic digital twin","One static video becomes an interactive 3D world with movable parts","DRAWER reconstructs scenes and articulates objects from video alone","Video in, interactive twin out: DRAWER fuses geometry and appearance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000881,"raw_usage":{"total_tokens":3778,"prompt_tokens":887,"completion_tokens":2891,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2810}},"tokens_in":503,"tokens_out":2891,"duration_ms":20417,"temperature":1.0,"reasoning_tokens":2810,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:29:04.929393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of furniture with articulation mechanisms that are not visually obvious from a closed state, such as push-to-open drawers, one-way swing cabinets with hidden dampers, or bifold doors, run DRAWER on a static video, and check whether the predicted joint type and axis let the simulated opening follow the real motion within the paper's trajectory-EMD threshold.","supporting_citations":[{"cited_title":"Bakedsdf: Meshing neural sdfs for real-time view synthesis","cited_arxiv_id":null,"evidence_quote":"Provides the neural surface reconstruction method whose extracted mesh supplies accurate geometry and anchors the splats."},{"cited_title":"Understanding 3d object interaction from a single image","cited_arxiv_id":null,"evidence_quote":"Supplies per-object articulation and hinge predictions from static images that DRAWER validates and refines with 3D geometry."},{"cited_title":"Video2game: Real-time interactive realistic and browser-compatible environment from a single video","cited_arxiv_id":null,"evidence_quote":"Closest prior system reconstructing an interactable scene from a single video; DRAWER extends it with articulation and the dual representation."}],"review_version":1}