{"id":"4c7272da-7de3-46db-8946-09950141e463","arxiv_id":"2505.17966","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Single-view mesh reconstruction models, despite strong computer vision benchmarks, fail robotics-specific accuracy, collision, stability, occlusion, and latency requirements on YCB-Video and Aria Digital Twin datasets.","lead":"This paper tests 11 single-view 3D mesh reconstruction models against five requirements for robotics simulation: accuracy, collision-free geometry, stability, occlusion handling, and speed. On two real robotics datasets, nearly all models miss these requirements by large margins, meaning the gap between computer vision progress and robotics needs is currently too wide for instant digital twin creation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'not ready' verdict hinges on author-set thresholds (notably 2mm Chamfer) never tied to a downstream manipulation success criterion; a controlled noise study on ground-truth meshes would show whether that bar is actually load-bearing.","rationale":"I read the paper as an empirical benchmark whose central claim is a negative readiness verdict for single-view mesh reconstruction in robotic manipulation. The experimental measurements (Chamfer distance, grasp transfer, collisions, stability, occlusion, latency) are internally consistent and, if anything, err in favour of the evaluated models: ground-truth masks are supplied, objects and frames are manually selected, scale is normalized to ground truth before error computation, and slow optimisation-based methods are excluded. Those choices make the observed failures conservative and do not threaten the negative direction of the results. The load-bearing weak point is the mapping from measurements to the binary 'ready/not ready' conclusion. The desiderata thresholds in Section II are motivated qualitatively but not validated against any specific manipulation task success criterion. The grasp-transfer result is the closest thing to task-level validation, but a 50% per-grasp transfer rate is not interpreted against a required threshold, and no closed-loop task success is measured. The paper itself acknowledges in Section IV-D that permissible error is context-dependent. Thus the central claim is conditional on these thresholds being appropriate for the intended robotics use cases. This is exactly the reader's identified weakest assumption, so I agree with the reader's assessment. The appropriate verdict remains CONDITIONAL: the empirical content appears sound and reproducible in principle, but the headline conclusion should be accompanied by task-validated tolerances and released evaluation code/data before being accepted as a general statement about robotics readiness. No concern about author conduct or internal consistency arose; the issue is the strength of the inference from measured errors to the global readiness verdict.","tokens_in":31964,"tokens_out":4873,"duration_ms":49664,"concrete_test":"Use the YCB-Video ground-truth meshes and the same MetaGraspNet antipodal grasp protocol: apply controlled isotropic surface perturbations (RMS 0, 2, 5, 10 mm) to GT meshes, recompute Chamfer distances, and measure both per-grasp transfer success and end-to-end pick-and-place success in PyBullet for a representative task (e.g., 10 YCB objects, 50 episodes each). If 5mm-perturbed meshes achieve task success comparable to unperturbed GT (within 5 percentage points), Desideratum 1's 2mm threshold is not load-bearing and the 'not ready' claim needs to be re-scoped to the specific thresholds and tasks tested.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II sets Desideratum 1 to 2mm Chamfer distance, citing 1mm grasp and sub-millimetre assembly references, but it never derives an acceptable task-level error from a concrete manipulation objective. The paper's own limitation section (IV-D) concedes that permissible reconstruction error is highly context-dependent. The headline conclusion that 'existing approaches fail to meet robotics-specific requirements' therefore depends on whether 2mm accuracy (and the analogous 5-degree stability and 2-second latency bars) are truly necessary for readiness. The grasp-transfer experiment in Fig. 3 is the only task-level evidence, yet it reports per-grasp transfer success of roughly 50% without specifying a required success rate or evaluating closed-loop task success; a robust manipulation planner can often compensate for imperfect individual grasp hypotheses by sampling and filtering. If a representative pick-and-place task tolerates 5mm surface error, the same reconstructions that fail the 2mm bar could still be 'ready' for that task. The central negative claim is thus stronger than the evidence directly supports, although the raw measurements are reported transparently and the manual object/frame selection likely biases in favour of the models, which makes the observed failures conservative.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether current single-view, category-agnostic mesh reconstruction models are ready to serve as the perception front-end for robot manipulation through instant digital twin creation. It proposes five robotics desiderata—reconstruction accuracy within 2 mm Chamfer distance, collision-free geometry, physical stability within 5° of the scene pose, bounded occlusion error, and scene reconstruction within 2 seconds—and evaluates eleven single-view reconstruction models on the YCB-Video and Aria Digital Twin datasets. The evaluation measures Chamfer distance, mesh collisions, object stability in PyBullet, occlusion resilience, latency, and memory, and adds an external grasp-transfer test in simulation. The main finding is that current models fail these desiderata by large margins, with typical Chamfer errors around 5 mm on YCB-Video and worse on Aria, frequent collisions and instabilities, substantially degraded occlusion handling, and inference times that exceed the 2 s target for most models.","tokens_in":32179,"tokens_out":6680,"duration_ms":78010,"significance":"If the empirical results hold, this is a valuable and timely negative result for the real-to-sim community: it documents, across two datasets, eleven models, and multiple complementary metrics, that single-view reconstruction performance on computer vision benchmarks does not transfer to robotics-grade physical simulation. The grasp-transfer experiment is a particular strength because it is an external, task-relevant validation against ground-truth meshes rather than a re-fit evaluation metric. The evaluation is also conservative in several respects—giving ground-truth masks to segmentation-dependent methods and manually selecting frames likely favors the models, so the observed failures are not easily explained away by pipeline artifacts. The main weakness is that the headline 'not ready' conclusion is stronger than the evidence directly supports, because the threshold values (notably 2 mm Chamfer) are asserted rather than derived from a downstream manipulation success criterion, a point the authors themselves concede in Section IV-D.","major_comments":[{"comment":"The thresholds of 2 mm Chamfer distance, 5° stability, and 2 s latency are presented as 'robotics-specific requirements', but they are not derived from a concrete manipulation objective. Section IV-D explicitly concedes that permissible reconstruction error is highly context-dependent and that deriving a systematic error-tolerance relationship is left for future work. The headline claim that 'existing approaches fail to meet robotics-specific requirements' is therefore stronger than the measurements alone establish: a task tolerating 5 mm surface error, or a pipeline using a robust grasp sampler, might accept the very reconstructions reported here. Please either add a controlled study that ties the thresholds to task outcomes (for example, corrupting ground-truth meshes to controlled Chamfer levels and measuring grasp-transfer or pick-and-place success), or rephrase the central claim as failure against the proposed desiderata rather than an unconditional statement about robotics readiness.","section":"§II, Desiderata 1–5; §IV-D"},{"comment":"The grasp-transfer experiment reports per-grasp transfer success rates near 50% without stating a required success rate, without a ground-truth-to-ground-truth baseline, and without an end-to-end pick-and-place evaluation. Since a manipulation planner can sample, rank, and filter many candidate grasps, a 50% per-grasp transfer rate is not by itself sufficient evidence that the reconstructions are unusable for manipulation. Please either report a task-level success metric with an explicit acceptance threshold and a baseline (e.g., grasps computed and evaluated on ground-truth meshes), or soften the conclusion drawn from this figure.","section":"§IV-C, Fig. 3"},{"comment":"The evaluation uses manually selected objects and frames and provides ground-truth masks to all segmentation-dependent methods. Providing ground-truth masks is a conservative choice that strengthens the negative result, but the manual selection protocol is not quantified: the paper does not report the distribution of object categories, view-points, occlusion levels, or pose diversity, nor does it analyze whether the chosen frames are representative of a deployment distribution. Please document the selection protocol in detail and, if possible, release the full object and frame lists so that other researchers can reproduce or extend the benchmark.","section":"§IV-A and §IV-B"}],"minor_comments":[{"comment":"InstantMesh is cited as [123] in the model description, but reference [123] is Instant3D; the rest of the paper and Figures 2, 3, 4, etc. cite InstantMesh as [27]. Please reconcile the citation.","section":"§IV-A"},{"comment":"There are several typos: 'This is ensures physical stability' in Desideratum 2 should be 'This ensures physical stability', and the abstract uses 'quantitively' instead of 'quantitatively'.","section":"§II, abstract"},{"comment":"The scale-estimation procedure based on ratios of principal-component standard deviations should include a caveat for objects with near-degenerate principal components (e.g., flat or axially symmetric objects), where the median ratio may be numerically unstable.","section":"§IV-B"},{"comment":"The legend and axis of Figure 10 both repeat 'Relative Error Increase (%)'; please clean up the caption and axis labels.","section":"Fig. 10"},{"comment":"The authors do not state whether evaluation code, selected frames, or model configuration files will be released; providing these would substantially increase the benchmark's reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a solid empirical benchmark study that could be a good fit for a robotics venue that publishes evaluations. The main revision needed is to align the title and abstract with the conditional nature of the findings: the paper's own limitation section already concedes that permissible error is task-dependent, so the unconditional 'not ready' framing overstates the evidence. The threshold concern is not, in my view, grounds for rejection, because the raw measurements are reported transparently and the evaluation is conservative in its use of ground-truth masks and manual frame selection; however, a controlled threshold-sensitivity study or a clearly hedged central claim is necessary before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best to tell you up front: this is the first systematic benchmark of single-view, category-agnostic mesh reconstruction models against robotics-specific requirements, and the news is bad for the models. Across 11 models and two datasets, reconstructions miss the 2mm Chamfer target by roughly 2-3x, collide in most scenes, are unstable in the majority of cases, degrade 40-95% under occlusion, and take seconds to tens of seconds per object. The grasp transfer experiment is the clincher: only about half of grasps computed on reconstructions transfer to ground-truth meshes.\n\nWhat the paper does well: the desiderata framework is sensible and overdue; the literature review is thorough and correctly distinguishes this work from multi-view real-to-sim pipelines; giving ground-truth masks to all methods isolates reconstruction quality; and manual object/frame selection likely biases in favor of the models, so the observed failures are conservative rather than inflated.\n\nThe soft spots are real but not load-bearing. The 2mm, 5-degree, and 2-second thresholds are author-set and not derived from a concrete manipulation task. The paper's own limitation section admits that permissible error is highly context-dependent. That means the 'not ready' verdict is stronger than the evidence strictly proves—a pick-and-place task tolerant of 5mm error might get by with some of these reconstructions. But the qualitative failures (majority collisions, large occlusion degradation) do not depend on the thresholds, and the grasp-transfer rate is far below any plausible task requirement. The occlusion evaluation uses a manual post hoc symmetry-flip correction, which is a minor blemish. No code or selected data are released, so replication is currently impossible.\n\nOverall: the central negative conclusion holds. The paper is honest, the data are reported transparently, and the analysis is careful. It deserves a serious referee. I would recommend acceptance with revision: release the evaluation code and object/frame selections, tie or bracket the thresholds to task-level success criteria, and make the occlusion correction more principled. Robotics practitioners and reconstruction researchers both get value from this.","headline":"A careful, honest benchmark showing single-view mesh reconstruction is not ready for physics-sim manipulation; the thresholds are debatable but the negative trend is robust.","tokens_in":32709,"tokens_out":2296,"would_cite":true,"duration_ms":20138,"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":"Current single-view 3D mesh models fail robotics-specific requirements for digital twin creation, an empirical evaluation on real robotics data finds.","keywords":["single-view 3D reconstruction","real-to-sim","digital twin","robotic manipulation","mesh reconstruction","physics simulation","object stability","occlusion handling"],"falsifier":"Run the same eight object-level models on a task whose success criterion is 5 mm surface distance instead of 2 mm; if grasp transfer to ground-truth meshes then exceeds roughly 90 percent, the paper's not-ready verdict would not generalise to that task, while success near 50 percent would confirm that reconstruction accuracy is the load-bearing constraint.","tokens_in":31764,"feed_emoji":"🤖","tokens_out":7251,"duration_ms":63290,"temperature":0.7,"pith_summary":"This paper asks whether today's single-view, category-agnostic mesh reconstruction models can serve as the perception front-end for instant digital twin creation in robotic manipulation: mapping one RGB(-D) image of a tabletop scene into complete object meshes ready for a physics simulator. To answer it, the authors define five robotics-specific desiderata, including surface accuracy within 2 mm Chamfer distance, collision-free geometry, stability within 5 degrees of the scene pose, occlusion error no more than 10 percent above visible-region error, and full-scene reconstruction within 2 seconds. They then evaluate eight object-level and three scene-level models on YCB-Video and Aria Digital Twin, two real robotics datasets with ground-truth meshes. Most models sit about 5 mm from the true surface on YCB-Video and about 1 cm on Aria, only about half of computed grasps transfer to the target meshes, and most reconstructed objects lack stable poses near their observed scene pose. The paper concludes that, despite strong computer vision benchmark results, existing approaches do not yet meet robotics-specific requirements for physics-based planning.","feed_headline":"Single-view 3D meshes fail robotics' accuracy bar","feed_subtitle":"Even the best models sit 5 mm off true surfaces and only half of computed grasps transfer to real meshes.","key_machinery":"The evaluative machinery combines five desiderata with a physics-aware measurement protocol on two annotated datasets. Reconstruction accuracy is measured as bidirectional Chamfer distance between 10,000 sampled surface points after ICP alignment seeded with 512 quaternion initialisations; grasp transfer is tested with antipodal grasp sampling from a parallel-jaw gripper, checking collisions, a 22.5-degree surface-normal alignment bound, and contact retention after shaking in simulation. Collisions are detected with FCL on objects placed at their estimated scene poses; stability is tested by perturbing each object 5 degrees away from candidate stable poses in PyBullet and checking whether it returns; occlusion handling compares Chamfer distance on visible versus occluded surface regions using a mask-recomputed ICP. The thresholds of 2 mm, zero collisions, 5 degrees stability, 10 percent occlusion degradation, and 2 seconds per scene are the yardstick that turns raw reconstruction error into a robotics-readiness verdict.","core_discovery":"The central claim is that single-view, category-agnostic full-mesh reconstruction models, used out-of-the-box on realistic robotics inputs, fail the requirements a digital twin for manipulation imposes. On YCB-Video, reconstructed surfaces are typically 5 mm from the closest ground-truth surface against a 2 mm target, and on Aria Digital Twin the median error is roughly twice as large. Grasp poses computed on the reconstructions transfer to the ground-truth meshes only about half the time, collisions occur in a majority of scenes for single-object models, and both object- and scene-level reconstructions are mostly unstable within 5 degrees of their observed pose. Occluded object regions raise Chamfer error by 40 to 95 percent for object-level models, while scene-level models that inpaint or jointly denoise objects stay near the 10 percent bound. Only SF3D and ZeroShape reconstruct a single object within roughly one second, and scene-level models take an order of magnitude longer, so the 2-second per scene target is met by none of them.","pith_inferences":["The 2 mm surface-distance threshold is not derived from a specific manipulation task; a task with 5 mm clearance, such as suction grasping of large objects, could succeed with current models even though the paper's headline verdict says not ready.","Grasp transfer rates depend on the chosen gripper and evaluation thresholds, so re-running the transfer test with a compliant or suction gripper could change the ranking and could raise success rates above the reported 50 percent.","A direct test of the paper's proposed remedies would be to fine-tune a fast model such as SF3D with physics-informed stability training while feeding it ground-truth depth, then re-running the same five-desiderata protocol to measure how much of the gap closes."],"forward_implications":["If correct, single-view reconstruction cannot currently serve as the perception front-end for real-time, physics-based manipulation planning at the tolerances the authors specify.","Practitioners should prefer scene-level models over single-object models when objects are physically close, because single-object models put the reconstructed meshes in mutual collision in a majority of evaluated scenes.","Occlusion handling improves sharply when models use scene context, either image inpainting before reconstruction or joint multi-object denoising, suggesting a concrete design direction for future reconstruction models.","Computational cost is the one desideratum with clear winners, with SF3D and ZeroShape producing objects in about 0.5 and 1 second, respectively, yet even they cannot handle a multi-object scene within the 2-second target."],"supporting_citations":[{"why":"Supplies the YCB-Video dataset with ground-truth meshes, poses, and masks used for the tabletop evaluation.","marker":"[38]"},{"why":"Supplies the Aria Digital Twin household scenes used to test generalization beyond canonical object arrangements.","marker":"[39]"},{"why":"SF3D, one of two object-level models achieving the best surface accuracy and the only model reconstructing an object in about 0.5 seconds.","marker":"[80]"},{"why":"DSO, the other best-accuracy model, fine-tuned with simulation feedback for physical stability at a 20-degree threshold.","marker":"[190]"},{"why":"One2345, a representative multi-view diffusion object model with high memory and runtime costs.","marker":"[26]"},{"why":"InstantMesh, a sparse-view large reconstruction model whose outputs also power the PhysGen3D scene pipeline.","marker":"[27]"},{"why":"Michelangelo, the model with the highest fraction of stable poses on YCB-Video.","marker":"[28]"},{"why":"ZeroShape, a fast feed-forward model producing reconstructions in about 1 second.","marker":"[142]"},{"why":"Supplies the antipodal grasp sampling and transfer evaluation procedure used to measure manipulation readiness.","marker":"[210]"},{"why":"PyBullet, the physics simulator used to test object stability under perturbation.","marker":"[41]"}],"fun_headline_variants":["Single-view mesh models fail robotics accuracy and speed","Reconstructions miss grasp, stability, and timing bars","One-shot 3D models fall short for robot digital twins","Mesh from single view fails 2mm and 2s robotics tests","Robotics-ready? Single-view meshes flunk key metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The verdict depends on the five thresholds the authors set in Section II—2 mm surface distance, zero collisions, stability within 5 degrees, at most 10 percent extra error on occluded regions, and 2 seconds per scene—which are presented as robotics requirements but are not derived from a specific manipulation task.","fun_headline_variants_meta":{"raw":{"variants":["Single-view mesh models fail robotics accuracy and speed","Reconstructions miss grasp, stability, and timing bars","One-shot 3D models fall short for robot digital twins","Mesh from single view fails 2mm and 2s robotics tests","Robotics-ready? Single-view meshes flunk key metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1477,"prompt_tokens":941,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":557,"tokens_out":536,"duration_ms":5781,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:36:56.974293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same eight object-level models on a task whose success criterion is 5 mm surface distance instead of 2 mm; if grasp transfer to ground-truth meshes then exceeds roughly 90 percent, the paper's not-ready verdict would not generalise to that task, while success near 50 percent would confirm that reconstruction accuracy is the load-bearing constraint.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DSO, the other best-accuracy model, fine-tuned with simulation feedback for physical stability at a 20-degree threshold."},{"cited_title":"MetaGraspNet: A Large-Scale Benchmark Dataset for Scene-Aware Ambidextrous Bin Picking via Physics-based Metaverse Synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the antipodal grasp sampling and transfer evaluation procedure used to measure manipulation readiness."}],"review_version":1}