{"id":"2338cd19-cc6a-4971-90be-c8105cae00af","arxiv_id":"2501.05483","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A decomposed VQ-VAE that generates diverse human grasps for rigid and deformable objects, with a Mesh UFormer backbone and normal-vector encoding for deformation simulation.","lead":"The paper introduces DVQ-VAE-2, a decomposed VQ-VAE that generates human grasps for rigid and deformable objects by encoding each finger and the palm into separate codebooks and decoding posture before position. It also adds a Mesh UFormer backbone with normal-vector-guided encoding to simulate object deformation, reporting small relative gains over existing methods on standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quality Index weight a=0.301 in Eq. 39 is fitted on benchmark test sets, and no scalar a can make DVQ-VAE-2 beat the best baseline on both GRAB and Obman; Table 1 therefore does not support the 'four benchmarks' superiority claim.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue: the Quality Index weight a = 0.301 is calibrated on Obman and GRAB, with GRAB being one of the four test sets, so the primary metric is test-set fitted. My independent check confirms and strengthens this concern: even setting calibration aside, Table 1 shows DVQ-VAE-2 is worse than GraspCVAE on Obman (2.89 vs 2.57), and the crossover thresholds derived from the four reported numbers show that no choice of a can make DVQ-VAE-2 superior on both GRAB and Obman. Thus the abstract's 'approximately 14.1% improvement across four benchmarks' is not supported by the table as presented. This does not mean the underlying system is without value: the raw penetration and displacement numbers, the decomposed VQ-VAE architecture, and the Mesh UFormer deformation modeling are reported in enough detail that the method can be evaluated on its own terms. The fix is not to abandon the work but to specify the metric calibration procedure, report leave-one-dataset-out weights, and state clearly which datasets contribute to the aggregate improvement. The reader's CONDITIONAL verdict is therefore the right one; my analysis does not move it to ACCEPT or REJECT, but it does make the requested revisions more specific and more urgent.","tokens_in":20352,"tokens_out":8807,"duration_ms":80898,"concrete_test":"Recompute Table 1's Quality Index under leave-one-dataset-out calibration: for each test set D in {HO-3D, FPHA, GRAB, Obman}, fit the weight a in Eq. (39) using only the other three datasets (or the Obman training split) following [51], then recompute relative improvements on D. Also report the crossover thresholds already implied by the table: GRAB requires a < 0.526, Obman requires a > 0.862. If no single a yields improvements on all four datasets, the 'across four benchmarks' headline must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim of roughly 14.1% relative improvement in grasp quality rests entirely on the Quality Index Q = a*x + (1-a)*y (Eq. 39, Sec. 5.2.1), with a = 0.301 computed using the method of [51] on the Obman and GRAB datasets. This is test-set calibration: GRAB is one of the four benchmark test sets in Table 1, so the evaluation metric itself is fitted to the test distribution, and rankings can be tuned by the choice of a. More concretely, even accepting the reported a = 0.301, Table 1 fails to show superiority across all four benchmarks: on Obman, GraspCVAE has Quality Index 2.57 while DVQ-VAE-2 has 2.89, i.e., worse. Crossover analysis makes the problem explicit. On GRAB, DVQ-VAE-2 beats GraspCVAE only when a < 0.526; on Obman, it beats GraspCVAE only when a > 0.862. Hence no single scalar weight a in Eq. 39 can make DVQ-VAE-2 better than the best baseline on both GRAB and Obman simultaneously. The claimed 14.1% aggregate therefore depends on either omitting Obman or averaging a loss on Obman against gains elsewhere, and it is computed with a metric weight calibrated on test sets. The raw penetration and displacement numbers are reported, but the headline metric and the 'four benchmarks' claim need to be reformulated before the superiority claim is accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DVQ-VAE-2, a decomposed VQ-VAE for generating human grasps on rigid and deformable objects. The hand is encoded as six separate components (five fingers and a palm), each with its own codebook, and decoding is split into a posture stage and a position stage. For deformable objects, the authors introduce a Mesh UFormer backbone that operates on voxelized mesh hierarchies, together with a normal-vector-guided position encoding and a hand-object contact quantification scheme that feeds penetration distances into the deformation network. Experiments compare against GraspCVAE, GraspTTA, ContactGen, GraspingField, and PointTransformerV3 on Obman, HO-3D, FPHA, GRAB, and HMDO, with the central claim being a roughly 14.1% relative improvement in grasp quality over state-of-the-art methods across four benchmarks.","tokens_in":20787,"tokens_out":4911,"duration_ms":44162,"significance":"The architectural ideas are plausible and potentially useful: part-aware discrete latent codes for hands, a dual-stage decoding strategy, and a mesh-based hierarchical backbone for deformation are all reasonable directions, and the paper reports extensive ablations. A clear strength is that raw penetration volume and grasp displacement numbers are reported for each method, and the authors state that source code and models will be released, which supports reproducibility. However, the headline evaluation rests on a new Quality Index whose weighting parameter is calibrated using test-set statistics, and the reported improvement numbers are internally inconsistent. Until the evaluation metric is fixed or the claims are reformulated to rest on the raw metrics, the central superiority claim is not supported.","major_comments":[{"comment":"The headline Quality Index is calibrated on test sets. The weight a=0.301 is computed on the Obman and GRAB datasets, while GRAB is one of the four test sets in Table 1, so the primary comparison metric is partly fit to the evaluation distribution. The problem is not merely formal: using the reported values, DVQ-VAE-2 beats GraspCVAE on GRAB only when a<0.526 and on Obman only when a>0.862, so no single scalar a can make the method better than the best baseline on both datasets simultaneously. Table 1 itself shows that with a=0.301, DVQ-VAE-2 has Quality Index 2.89 on Obman versus GraspCVAE's 2.57. The 'four benchmarks' superiority claim and the 14.1% aggregate therefore need to be reformulated; the raw penetration and displacement numbers should be the primary evidence, or the authors should report robustness of the ranking over a range of a values, or fix a without using test-set statistics.","section":"§5.2.1, Eq. (39), Table 1"},{"comment":"The headline improvement numbers are internally inconsistent. The abstract claims 'approximately 14.1%' relative improvement in grasp quality across four benchmarks, but §5.5 reports 13.3% on HO-3D, 13.7% on FPHA, and 1.2% on GRAB for DVQ-VAE, and then reports for DVQ-VAE-2 a 5.86% to 20.43% improvement over PointNet, not over state-of-the-art methods. No calculation in the paper derives 14.1% from Table 1, and the comparison basis changes between 'state-of-the-art' and 'PointNet backbone.' Please clarify what exactly is averaged and against which baseline.","section":"Abstract and §5.5, Table 1"},{"comment":"The ablation table contains a duplicated row label with conflicting numbers: 'DVQ-VAE + Dual-Stage (Two Encoders)' appears with Penetration 4.44, Grasp Disp 3.61, Quality Index 3.86, and again with Penetration 5.36, Grasp Disp 2.75, Quality Index 3.54. This makes it impossible to infer which configuration corresponds to the reported results and undermines the ablation claims for the dual-stage strategy. Please assign unique names to each variant and reconcile the numbers.","section":"Table 4"}],"minor_comments":[{"comment":"The negative weight λm=-50 on Lm is never explained. Since Lm as defined in Eq. (14) is a similarity (larger overlap is better), a negative coefficient is mathematically coherent if the loss is to be minimized; however, the text calls it a 'contact map consistency loss' and lists it under losses to improve interaction. Please state the sign convention explicitly and verify that the total loss is bounded in a sensible range.","section":"§3.5, Eq. (17), §5.4"},{"comment":"The model name is written as 'D-VQVAE-2' in one place ('After integrating Mesh UFormer, we propose D-VQVAE-2'); please make the nomenclature consistent throughout.","section":"§5.5"},{"comment":"The first row labeled 'Originality' has no method name and reports Chamfer Distance 0.00; it should be identified as the un-deformed input mesh or removed, since it is not a compared method.","section":"Table 2"},{"comment":"Table 1 does not state that ContactGen is trained on GRAB while the other methods are trained on Obman; this makes the GRAB comparison in-domain for ContactGen, which should be noted when interpreting the results.","section":"Table 1 and §5.3"},{"comment":"The Quality Index is introduced with 'lower is better,' but the text does not define what x and y are immediately before Eq. (39); the reader has to infer from the sentence that x is penetration volume and y is displacement. Please define both symbols explicitly.","section":"§5.2.1"}],"recommendation":"major_revision","confidential_remarks":"The core architecture is a clear extension of the authors' ECCV 2024 DVQ-VAE paper, with the deformable-object component being the main new contribution. The evaluation issues are the primary obstacle: the Quality Index weight is calibrated on test-set data, and the aggregate improvement claim is not reproducible from the tables. These problems are fixable within the manuscript's scope by reformulating the claims and reporting robustness analyses, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an honest extension of the authors' ECCV 2024 DVQ-VAE, and the deformable-object part is genuinely new. The Mesh UFormer backbone, the normal-vector-guided position encoding, and the bidirectional penetration quantification are real contributions, and the HMDO experiments show consistent gains over PointNet, PTV3, and Deformation Graph on contact distance and Chamfer distance. The dual-stage decoding idea is also reasonable.\n\nThe problem is the headline claim. The Quality Index Q = a*x + (1-a)*y is defined with a = 0.301 computed on Obman and GRAB, and GRAB is one of the four test benchmarks. The abstract says 'four benchmarks', the introduction says 'five', and the results section reports 13.3%, 13.7%, and 1.2% on three datasets before mentioning 5.86%–20.43% for the backbone. No single scalar a makes DVQ-VAE-2 beat the best baseline on both GRAB and Obman: on GRAB it beats GraspCVAE only for a < 0.526, on Obman only for a > 0.862. The reported a = 0.301 makes it win on GRAB but lose on Obman (Quality Index 2.89 vs 2.57 for GraspCVAE). So '14.1% across four benchmarks' is at best an average of wins and losses using a metric fitted to test sets. That is a load-bearing flaw for the superiority claim.\n\nThere are also internal inconsistencies that need fixing: the abstract's 14.1% does not match the numbers in Sec. 5.5; Table 4 lists 'DVQ-VAE + Dual-Stage (Two Encoders)' twice with different numbers, and the DVQ-VAE-2 row there (Quality Index 4.28) does not match Table 1's HO-3D value (2.95); and the contact-map consistency loss weight is lambda_m = -50, which with Lm defined as an overlap ratio in Eq. 14 is the wrong sign unless there is a typo.\n\nOn the plus side, the paper ships code, the deformable-object experiments are detailed, and the human evaluation is run with 20 participants. The deformable contribution is solid enough that I would not desk-reject it, but the authors need to reset the comparison metric, fix the inconsistent tables, and report Pareto-style trade-offs between penetration and displacement instead of a single calibrated index. The raw numbers support modest gains on some datasets, not a consistent superiority claim.\n\nMy recommendation: send it to peer review, but with a clear instruction that the metric calibration and the inconsistent tables must be addressed before the claim is credible. A good referee can separate the real deformable-object contribution from the overclaimed headline.","headline":"The deformable-object components are genuinely new and the HMDO gains look real, but the Quality Index is calibrated on test benchmarks and the 'four benchmarks' superiority claim does not survive a crossover analysis.","tokens_in":21250,"tokens_out":3974,"would_cite":false,"duration_ms":34937,"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":"Decomposing the hand into fingers and palm, and decoding posture before position, produces higher-quality grasps and more accurate deformation simulation, with about 14.1% average improvement over prior methods.","keywords":["grasp generation","VQ-VAE","decomposed architecture","hand-object interaction","deformable objects","mesh transformer","MANO hand model","normal vector position encoding"],"falsifier":"Recompute the quality index using a weight learned only from the training split of Obman (or set to a=0.5) and re-run the comparisons on HO-3D, FPHA, and GRAB; if the relative improvement over GraspTTA and ContactGen disappears or reverses, the central claim fails. Alternatively, check whether human preference ratings or physical simulation success track the quality index when the weight is varied — if small changes in a flip the method ranking, the claimed improvement is metric-dependent.","tokens_in":20153,"feed_emoji":"🖐️","tokens_out":9217,"duration_ms":82518,"temperature":0.7,"pith_summary":"Humans grasp objects with a hand made of semi-independent parts, yet most grasp models encode the hand as one whole and predict its pose in a single step. This paper argues that decomposing the hand into five fingers and a palm, giving each its own discrete codebook, and then generating the grasp in two stages—posture first, position second—produces more realistic and more varied grasps. For deformable objects, the proposal adds a mesh-understanding network that uses surface normal directions to compute where a hand presses into the object and to simulate the resulting deformation. The authors report that this design raises grasp quality by about 14.1% on average over previous methods across four benchmarks, improves hand-object contact distance on deformable objects, and increases grasp diversity. If correct, the paper shows that respecting the part structure of the hand and the order of grasp decisions matters more than a single bigger network.","feed_headline":"Part-aware hand encoding improves grasp quality by 14%","feed_subtitle":"Encoding fingers and palm separately yields more natural grasps and better deformation modeling.","key_machinery":"The load-bearing mechanism is the decomposed, multi-codebook VQ-VAE: the hand's five fingers and palm are each quantized against their own codebook, and an autoregressive model generates the index sequence conditioned on the object codebook, preserving the discrete, part-determined structure of grasps. Around this core sit the dual-stage decoder (posture parameters via MANO under skeletal angle gating, then position parameters) and, for deformable objects, Mesh UFormer, a symmetric mesh encoder–decoder that downsamples the mesh by voxel clustering, propagates features between resolutions through KD-tree point mappings, and uses a normal-vector-guided position encoding inside its attention. The hand-object contact quantification completes the mechanism: it computes both hand-to-object and object-to-hand penetration distances using normal-vector dot products and ray casting, and those distances are fed to the deformation network.","core_discovery":"The central claim is that a part-aware discrete latent representation, paired with an ordered two-stage decoder, yields more realistic and more adaptable grasps. Concretely, DVQ-VAE-2 divides the hand into five fingers and a palm, maps each part to its own VQ-VAE codebook (a learned discrete dictionary), and encodes the object with two separate encoders, one for grasp type and one for grasp pose. Decoding is sequential: a PixelCNN autoregressively predicts the hand codebook indices conditioned on the object codebook, then a posture decoder produces MANO posture parameters under skeletal joint-angle gating, and a position decoder then produces the remaining position parameters. For deformable objects, Mesh UFormer replaces PointNet as the backbone, using voxel-based hierarchical mesh sampling and normal-vector-guided position encoding to predict per-vertex deformation from bidirectional hand-object penetration distances. The authors report improvements in Quality Index of 13.3%, 13.7%, and 1.2% on HO-3D, FPHA, and GRAB with the PointNet backbone, and further gains of 5.86% to 20.43% in Quality Index over PointNet on rigid datasets and 2.23% in Hand-object Contact Distance over PointTransformerV3 on the HMDO dataset.","pith_inferences":["A direct robustness check suggested by the metric design: recalibrate the quality index weight using only the training partition of Obman and see whether the 14.1% average improvement persists; if it shrinks, the headline number is partly test-set dependent, even though raw penetration and displacement improvements may remain.","The part-level codebooks could transfer to robot hands with different morphologies: treating each finger as a conditionable module would let the same framework generate grasps for hands with varying finger lengths or counts by swapping or adding codebooks.","The normal-vector-guided position encoding is a general inductive bias for contact-driven mesh deformation, so it could be applied to other soft-body manipulation problems where local surface orientation determines how the material yields."],"forward_implications":["Grasp generation becomes much faster: DVQ-VAE-2 generates a batch of grasps in 0.11 s on a single GPU, compared with roughly 19–20 s for GraspTTA and over 100 s for ContactGen, while matching or exceeding their contact ratios.","The part-aware codebooks increase grasp diversity: cluster size on HO-3D rises from about 0.81 for GraspCVAE/GraspTTA to roughly 3.86, so a single object yields a wider range of natural postures without test-time adaptation.","For deformable objects, deformation accuracy improves specifically at hand-contact regions (2.23% better Hand-object Contact Distance than PointTransformerV3) and unwanted far-field deformation drops sharply (a 68.82% relative improvement in Chamfer Distance), using only about 14.56% of PointTransformerV3's parameters.","The dual-stage decoding strategy alone improves the Quality Index by 45.2% for DVQ-VAE and 47.6% for VQ-VAE in the ablations, implying that the ordering of posture-then-position is itself a substantial source of the gain.","The learned position decoder can refine grasps from other methods: applying it to grasps produced by other models reduces penetration, as shown in Figure 10."],"supporting_citations":[{"why":"Provides the primary baseline GraspTTA/CVAE method with test-time adaptation and object-centric contact loss that DVQ-VAE-2 is compared against on rigid benchmarks.","marker":"[1]"},{"why":"ContactGen baseline; the CVAE contact-map grasp generation method whose quality index and diversity numbers DVQ-VAE-2 must beat.","marker":"[16]"},{"why":"Grasping Field baseline using implicit VAE representations; comparison target for penetration, displacement, and quality index.","marker":"[2]"},{"why":"The VQ-VAE method that provides the discrete latent codebook framework which the paper extends to part-aware decomposition.","marker":"[20]"},{"why":"MANO parametric hand model used to decode posture and position parameters into a 778-vertex hand mesh.","marker":"[21]"},{"why":"Obman dataset used to train the rigid-object grasp models and to calibrate the quality index weight.","marker":"[11]"},{"why":"HMDO dataset of hand interactions with deformable objects, used to train and test the deformation simulation network.","marker":"[14]"},{"why":"PointTransformerV3 baseline for deformable-object deformation simulation; the strongest comparison in hand-object contact distance.","marker":"[23]"},{"why":"Deformation Graph method for shape manipulation; used as a deformable-object baseline.","marker":"[39]"},{"why":"Adaptive weight learning method used to set the a=0.301 balance in the quality index, which the headline improvement depends on.","marker":"[51]"}],"fun_headline_variants":["Decomposed hand encoding lifts grasp quality by 14%","Part-aware VQ-VAE improves grasping realism by 14%","Two-stage decoder yields more natural grasps","Discrete hand parts for better object grasping","Hand decomposition boosts grasp accuracy and adaptability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's headline result depends on the assumption that the single weight balancing penetration volume against grasp displacement in its quality index, chosen using the Obman and GRAB datasets, is a stable property of grasp quality across all four benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["Decomposed hand encoding lifts grasp quality by 14%","Part-aware VQ-VAE improves grasping realism by 14%","Two-stage decoder yields more natural grasps","Discrete hand parts for better object grasping","Hand decomposition boosts grasp accuracy and adaptability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3103,"prompt_tokens":1088,"completion_tokens":2015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":1941}},"tokens_in":704,"tokens_out":2015,"duration_ms":14965,"temperature":1.0,"reasoning_tokens":1941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:37:04.638125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the quality index using a weight learned only from the training split of Obman (or set to a=0.5) and re-run the comparisons on HO-3D, FPHA, and GRAB; if the relative improvement over GraspTTA and ContactGen disappears or reverses, the central claim fails. Alternatively, check whether human preference ratings or physical simulation success track the quality index when the weight is varied — if small changes in a flip the method ranking, the claimed improvement is metric-dependent.","supporting_citations":[{"cited_title":"Hand-object contact con- sistency reasoning for human grasps generation,","cited_arxiv_id":null,"evidence_quote":"Provides the primary baseline GraspTTA/CVAE method with test-time adaptation and object-centric contact loss that DVQ-VAE-2 is compared against on rigid benchmarks."},{"cited_title":"Contactgen: Generative contact modeling for grasp generation,","cited_arxiv_id":null,"evidence_quote":"ContactGen baseline; the CVAE contact-map grasp generation method whose quality index and diversity numbers DVQ-VAE-2 must beat."},{"cited_title":"Grasping field: Learning implicit representations for human grasps,","cited_arxiv_id":null,"evidence_quote":"Grasping Field baseline using implicit VAE representations; comparison target for penetration, displacement, and quality index."},{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"The VQ-VAE method that provides the discrete latent codebook framework which the paper extends to part-aware decomposition."},{"cited_title":"Learning joint reconstruction of hands and manipulated objects,","cited_arxiv_id":null,"evidence_quote":"Obman dataset used to train the rigid-object grasp models and to calibrate the quality index weight."},{"cited_title":"Hmdo: Markerless multi-view hand manipulation capture with deformable objects,","cited_arxiv_id":null,"evidence_quote":"HMDO dataset of hand interactions with deformable objects, used to train and test the deformation simulation network."},{"cited_title":"Embedded deformation for shape manipulation,","cited_arxiv_id":null,"evidence_quote":"Deformation Graph method for shape manipulation; used as a deformable-object baseline."},{"cited_title":"Adaptive Weight Learning for Multiple Outcome Optimization With Continuous Treatment","cited_arxiv_id":"2402.11092","evidence_quote":"Adaptive weight learning method used to set the a=0.301 balance in the quality index, which the headline improvement depends on."}],"review_version":1}