{"id":"3913be18-5113-44cf-8d69-6c4db4f8b3d3","arxiv_id":"2505.06832","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UniDiffGrasp integrates VLM part identification with constrained grasp diffusion to achieve reported single-arm (0.876) and dual-arm (0.767) grasp success rates on a real robot.","lead":"This paper builds a robot grasping system that combines a vision-language model with a part-guided diffusion method, allowing a robot to grasp specific functional parts of objects in response to natural-language requests. It reports high success rates in single-arm and dual-arm tests on a real Baxter robot and argues that the integration is the key advance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GSR and CFR as defined do not verify that executed grasps contact the VLM-identified part; the headline part-grasping claim is therefore not measured.","rationale":"The reader's verdict (CONDITIONAL) is appropriate, and the reader's weakest_assumption focuses on the upstream reliability of VLM reasoning plus GroundedSAM/VLPart segmentation. I agree that this is unmeasured and important, but the single most load-bearing concern is downstream: even if Eqs. (2)-(3) always produced perfect masks, the reported GSR and CFR do not check whether the executed grasp is on the masked part. GSR is only lift-and-hold success; CFR is only collision avoidance against the object mesh. The dual-arm selection in Section III-C adds force-closure and maximal-distance criteria, but those are stability and balance criteria, not part-targeting criteria; a wrong-side pair of grasps can still pass them. Thus the paper's central claim -- open-vocabulary constrained grasping of specific functional parts -- is underdetermined by the presented evidence. The fix is straightforward and cheap: label each successful trial by whether the gripper contacts the target part, and report this alongside GSR. If the relabeling shows most successes are on the intended part, the concern is resolved and the method is supported. If not, the headline numbers must be scoped down to unconstrained object grasping. Given this, I keep the reader's CONDITIONAL verdict: the integration is plausible and the CGDF component has independent support, but the evaluation must be augmented before the abstract's 'significantly surpassing' claim can be accepted.","tokens_in":11048,"tokens_out":6658,"duration_ms":71694,"concrete_test":"Re-analyze the recorded trials behind Tables I and II: for every executed grasp, compute (a) IoU of the VLPart mask M*_p against a manual part annotation; (b) the distance from the selected gripper contact region to P_t; and (c) a binary 'part-targeted' label requiring both grippers to contact the masked part within a small tolerance. Report this Part-Targeting Rate (PTR) per object next to GSR. If PTR is close to GSR and mask IoU is high, the part-guided claim is substantiated. If PTR is substantially below GSR, the reported success rates overstate constrained, part-targeted grasping and the paper must be revised to report PTR and/or task-success metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that the headline metrics do not verify the central semantic constraint. Section IV defines GSR as 'successful physical grasp executions (lift and stable hold)' and CFR as 'no collision between the gripper model and the target object mesh'; neither requires the executed grasp to contact the VLM-identified part mask M*_p / point cloud P_t from Eq. (3). Consequently, a system that ignores the part constraint and produces any stable grasp on the object could still score high GSR for objects like mugs, pans, or kettles, since lifting those objects by the rim or body is physically possible. The paper's claim that UniDiffGrasp generates grasps 'accurately targeted to the VLM-identified functional part' (end of Section III-B) is therefore not actually measured. Tables I and II are compatible with unconstrained grasping; even if the VLM and segmentation stack in Eqs. (2)-(3) failed entirely, the reported GSR could remain high. This is more central than the reader's upstream segmentation concern: even with perfect masks, the current evaluation cannot distinguish a part-guided grasp from a generic one. Additionally, Figure 2's 'error messages (if exist)' input to the VLM hints at an unspecified retry loop, so the definition of an 'attempt' also needs clarification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniDiffGrasp, an end-to-end pipeline for open-vocabulary constrained grasping on a dual-arm Baxter robot. A VLM (GPT-4o) parses the user instruction and visual input into a target object, functional part, and operation mode (Eq. (1)); GroundedSAM and VLPart ground these into an object point cloud P and a part point cloud P_t (Eqs. (2)-(3)); a Constrained Grasp Diffusion Field (CGDF) with part-guided diffusion generates 6-DoF grasp candidates conditioned on P and P_t, from which a lowest-energy grasp is selected (Eq. (4)). For dual-arm tasks, two target regions are obtained by semantic or geometric splitting, per-arm part-guided diffusion generates grasp sets, and a selection procedure filters by energy threshold, inter-gripper collision, force closure, and maximal distance (Eqs. (7)-(9)). The paper reports real-world experiments on seven single-arm objects and four dual-arm objects, 30 trials per object-method condition, with average GSR 0.876 and 0.767 and CFR 0.900 and 0.850, respectively, compared against ThinkGrasp and a FPS+KNN baseline.","tokens_in":11258,"tokens_out":6021,"duration_ms":55472,"significance":"If the reported results are valid, the paper's main contribution is a working integration of VLM reasoning with a constrained diffusion grasp generator, showing that part masks from open-vocabulary segmentation can serve as geometric constraints for zero-shot constrained grasping, and that a simple pair-selection procedure can produce stable dual-arm grasps. The real-world deployment on a physical robot with 30 trials per condition is a concrete engineering effort, and using physical execution outcomes rather than only simulated metrics avoids circularity. However, the significance is weakened by the evaluation design: the headline metrics do not verify that the executed grasps actually contacted the intended part, and the statistical basis for the claimed superiority is absent. The novelty of the pipeline itself is largely integrative—each component (GPT-4o, GroundedSAM, VLPart, CGDF) is existing—so the contribution rests heavily on the strength of the experimental validation.","major_comments":[{"comment":"The central claim that UniDiffGrasp performs part-targeted grasping is not supported by the reported metrics. GSR is defined as successful physical grasp executions (lift and stable hold) and CFR as the percentage of generated grasp poses with no collision between the gripper model and the target object mesh; neither definition requires the executed grasp to contact the VLM-identified part mask M*_p or the target point cloud P_t from Eq. (3). For objects such as mugs, pans, and kettles, a grasp on the rim or body can lift the object successfully, so a system that ignores the part constraint entirely could still achieve high GSR and CFR. The sentence at the end of Section III-B saying the final grasp is accurately targeted to the VLM-identified functional part is therefore not measured by the current evaluation. I recommend adding a part-contact metric (e.g., verifying that the gripper contact region intersects M*_p in the 3D point cloud) and reporting it per object.","section":"IV (Tables I and II)"},{"comment":"All comparisons are based on point estimates from 30 trials per object-method condition, with no confidence intervals, variance measures, or significance tests. For example, the single-arm average GSR of 0.876 (Ours) versus 0.705 (ThinkGrasp) is reported as significantly surpassing, but with 30 trials per object the per-object binomial 95% confidence intervals are roughly plus or minus 0.17, and even pooled across seven objects the interval for 0.876 would be about plus or minus 0.045; no test accounts for object-to-object variability. The claim of significance is thus unsupported. Please report per-object confidence intervals or raw counts (e.g., 27 out of 30) and, if possible, a paired statistical test across objects (e.g., Wilcoxon signed-rank) or a mixed-effects model.","section":"IV-A and IV-B, Tables I and II"},{"comment":"The dual-arm baseline is not a state-of-the-art method and does not isolate the contributions of the proposed system. The baseline uses FPS+KNN to generate random target regions and omits semantic splitting, energy-threshold filtering, force-closure checking, and maximal-distance selection, so the comparison conflates target-region quality with the pair-selection procedure. A more informative comparison would include a baseline with FPS+KNN regions but the same selection pipeline, or a semantic-splitting-only variant, to attribute the reported GSR and CFR gains to specific components. The paper should also state the baseline's underlying grasp generator and give the exact K and distance parameters used.","section":"IV-B, Table II"},{"comment":"The pipeline's upstream segmentation step is a load-bearing assumption that is not measured. Equation (3) feeds the VLPart mask M*_p directly into P_t; if GroundedSAM or VLPart returns a wrong or geometrically incoherent mask, the part-guided diffusion cannot recover, yet the paper reports no segmentation success rate or analysis of failure cases. In addition, Fig. 2 shows error messages being sent back to the VLM, which implies an unspecified retry loop; the paper never defines whether a trial attempt includes such retries or whether failed segmentations are counted in the GSR denominator. Please specify the retry protocol and report the part-segmentation accuracy separately from grasp execution.","section":"III-A and Fig. 2"}],"minor_comments":[{"comment":"The text 'UniDiffGrasp, as detailed in Section IV-B and IV-A' is self-referential and should be corrected to point to the appropriate sections.","section":"IV-A"},{"comment":"The hyperparameters T (diffusion steps), N (candidate grasps per arm), the energy threshold delta, and the force-closure quality threshold are never given; please report them in the experimental section.","section":"III-B and III-C"},{"comment":"Equation (4) selects the final grasp by minimizing global energy E_theta at k=0, while the guided diffusion uses max(e'_k, e''_k); please clarify whether this selection could choose a grasp that is not on the target part and, if not, why the guided energy is not used for selection.","section":"III-B, Eq. (4)"},{"comment":"The CFR metric is computed against a ground-truth object mesh, but the collision-checking method (signed distance field, gripper model resolution, etc.) is not described; please add one sentence describing the checker.","section":"IV, CFR definition"},{"comment":"There are a few textual inconsistencies, such as 'Table 1' vs 'TABLE I' and the caption in Fig. 1 where 'BottleKnife' lacks a space; these should be cleaned up before publication.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main stress-test concern is justified: the evaluation as written could not distinguish part-guided from unconstrained grasping. This is an evaluation gap rather than a technical flaw in the method itself, and it can be fixed by adding part-contact verification and statistical rigor. The paper would also benefit from a stronger dual-arm baseline and component ablations. I do not see evidence of circularity; the success rates are physical execution outcomes independent of the energy model's internal predictions. The strongest version of this paper would be considerably more convincing, so I recommend major revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The system integration on display here is real. UniDiffGrasp takes VLM-derived part masks and feeds them directly as geometric constraints into CGDF's part-guided diffusion, then adds a sensible semantic/geometric splitting scheme for dual-arm coordination with force-closure-aware pair selection. The authors are honest that the core diffusion is CGDF's, and they have built a full pipeline and deployed it on a Baxter. That is a useful engineering contribution, especially the idea of using open-vocabulary segmentation to define P_t for constrained diffusion without retraining.\n\nThe soft spots, though, are load-bearing. The stress-test note is right: GSR and CFR as defined do not measure whether the executed grasp contacted the VLM-identified part. A stable lift of a mug by the rim or a pan by the body counts as a success, so even a system that entirely ignored the part constraint could score well. The paper's central claim—that UniDiffGrasp generates grasps 'accurately targeted' to the functional part—is therefore not actually verified by Tables I and II. That is more central than the upstream segmentation worry; even with perfect masks, the evaluation can't distinguish a part-guided grasp from a generic one.\n\nSecond, the statistical support is thin. Thirty trials per condition produce point estimates with no confidence intervals or significance tests. 'Significantly surpassing' is not supported. The dual-arm baseline is also very weak—random FPS+KNN region generation—so beating it is not a high bar. And the paper doesn't release code, data, or precise threshold values (energy delta, force-closure threshold, retry behavior), which limits independent verification. The 'error messages (if exist)' in Fig. 2 hints at a retry loop that is never specified, so the definition of an 'attempt' is unclear.\n\nOn the math and citations: the method itself is coherent and the reliance on CGDF is properly attributed. The contribution is an integration, not a new grasp model, and the paper mostly says so. The main gap is empirical, not theoretical.\n\nWho is this for? Roboticists working on language-conditioned part grasping and bimanual manipulation. They'll find the pipeline design useful even if the evaluation is incomplete. I'd send it to peer review because the system is real and the idea is worth refining, but it needs heavy revision: a part-contact verification metric (or manual labeling), statistical tests, stronger baselines, and ideally release of artifacts. As it stands, the evidence does not support the abstract's claim.","headline":"A plausible and genuinely integrated VLM-to-diffusion grasping pipeline, but the reported metrics don't verify that grasps actually contact the intended part, so the headline 'significantly surpassing' claim is under-supported.","tokens_in":11831,"tokens_out":1395,"would_cite":false,"duration_ms":15778,"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":"UniDiffGrasp claims that a VLM-chosen functional part, converted into a point-cloud constraint, directly guides diffusion-based grasp synthesis, reaching 0.876 single-arm and 0.767 dual-arm success on a real robot without retraining.","keywords":["open-vocabulary grasping","task-oriented grasping","part-guided diffusion","constrained grasp generation","dual-arm manipulation","vision-language models","affordance grounding","6-DoF grasp synthesis"],"falsifier":"Run the identical objects and instructions with ground-truth part masks substituted for the VLM-segmentation output; if success rates stay near the reported levels, the VLM-driven constraint stage is not doing the work, whereas a large drop would confirm that mask quality is the bottleneck. A stage-wise failure log that counts missed or malformed part masks separately from grasp-pose collisions would also settle where the pipeline breaks.","tokens_in":10818,"feed_emoji":"🦾","tokens_out":7095,"duration_ms":64852,"temperature":0.7,"pith_summary":"UniDiffGrasp is an attempt to close the gap between what a vision-language model understands about a task and where a robot actually places its grippers. The paper's claim is that the part a VLM identifies from a natural-language instruction — the handle, the keyboard, both handles — can be turned directly into a geometric constraint that steers a diffusion-based grasp generator, so no retraining or task-specific data is needed. On a physical dual-arm robot, the authors report 0.876 grasp success in single-arm trials and 0.767 in dual-arm trials, with the constrained diffusion producing higher collision-free rates than baselines that feed the same semantic labels into general-purpose grasp detectors. The broader motivation is that household and industrial tasks usually require grasping by a functional part, not by the object as a whole, and often require two arms working together.","feed_headline":"Part-guided diffusion lifts open-vocab grasping to 87.6%","feed_subtitle":"A VLM's part choice becomes a geometric constraint, letting one diffusion model grasp new parts and coordinate two arms without retraining.","key_machinery":"The load-bearing mechanism is Part-Guided Diffusion inside a Constrained Grasp Diffusion Field: a diffusion model trained only on unconstrained grasps is made to respect a target region by evaluating candidate poses against both the global object point cloud and the part point cloud, with guided energy $\\max(e'_k, e''_k)$ and an adaptive score that favors whichever constraint is harder to satisfy. The VLM-identified part becomes that part point cloud through hierarchical segmentation — an object mask from GroundedSAM followed by a part mask from VLPart — so language is effectively converted into geometry. This mechanism is what lets the pipeline skip retraining and still produce 6-DoF grasps concentrated on the requested functional region.","core_discovery":"The central discovery is that a semantic part label, once grounded as a point cloud through hierarchical open-vocabulary segmentation, works as a direct conditioning input for the Constrained Grasp Diffusion Field's Part-Guided Diffusion strategy. At each diffusion step, the energy model is evaluated against both the whole object and the target part; the effective guiding energy is the maximum of the two, so the score steers candidate grasps into the target region while keeping them globally stable and collision-free. This makes constrained grasping a zero-shot operation: the same unconstrained-trained generator adapts to arbitrary part constraints at inference time. For dual-arm tasks, UniDiffGrasp defines two target regions, either by splitting semantic parts or by geometric division, generates candidate grasps for each arm independently, then filters by energy, inter-gripper collision, and force closure, and finally selects the pair with maximal gripper separation. The reported outcome is grasp success of 0.876 in single-arm and 0.767 in dual-arm scenarios across seven single-arm and four dual-arm household object classes.","pith_inferences":["If upstream mask quality is the real bottleneck, then improving open-vocabulary part segmentation without touching the diffusion module should raise end-to-end success further; the paper does not test this directly.","The same interface would accept constraints from sources other than a VLM, such as a user clicking a point in the image or a separate affordance model, so the contribution generalizes beyond the specific language pipeline.","A direct test of this modularity would be swapping the VLM for a different part proposer and measuring whether grasp success tracks mask quality rather than the choice of language model.","The dual-arm selection criteria are generic enough to extend beyond rigid household objects to cooperative manipulation of larger or deformable items, though the paper's experiments do not demonstrate that extension."],"forward_implications":["A single diffusion model trained only on unconstrained grasps can be reused zero-shot for any new target part described in language, because the constraint enters only as a point cloud at inference time.","Natural-language task commands such as 'pour,' 'stew,' or 'type' become part-specific grasps without per-object retraining or constraint-specific datasets.","Dual-arm coordination reduces to defining two target regions and pairing independently generated per-arm grasps by collision, force closure, and separation, rather than learning a dedicated dual-arm policy.","The gap in collision-free rate between UniDiffGrasp and the baselines indicates that the constrained diffusion, not the semantic labels alone, supplies the geometric precision.","The reported gains are largest on objects with complex or small functional parts, such as pans, mugs, kettles, and pots, where post-filtering unconstrained grasps is least sample-efficient."],"supporting_citations":[{"why":"Supplies the Part-Guided Diffusion strategy inside a Constrained Grasp Diffusion Field, the core grasp-generation engine that consumes VLM-derived part constraints.","marker":"[11]"},{"why":"Supplies the vision-language model that interprets the user instruction and outputs target object, part, and operation mode.","marker":"[27]"},{"why":"GroundedSAM performs open-vocabulary object segmentation, producing the global object point cloud used as an input constraint.","marker":"[28]"},{"why":"VLPart performs open-vocabulary part segmentation on the object mask, producing the target part point cloud that conditions diffusion.","marker":"[22]"},{"why":"ThinkGrasp is the state-of-the-art baseline in single-arm experiments, representing VLM reasoning paired with a general-purpose grasp detector.","marker":"[7]"},{"why":"GraspNet is the general-purpose grasp generator used in the comparison baselines, illustrating the decoupled approach UniDiffGrasp claims to outperform.","marker":"[8]"},{"why":"AnyGrasp is another general-purpose grasp detector cited as the kind of downstream generator that can fail under precise VLM-imposed constraints.","marker":"[9]"},{"why":"Provides the differentiable force closure estimator used in dual-arm pair selection to discard unstable cooperative grasps.","marker":"[29]"}],"fun_headline_variants":["Zero-shot dual-arm grasping via part-guided diffusion","Part labels guide diffusion for 87.6% grasp success","VLM+diffusion: open-vocab grasps with dual-arm coordination","Diffusion adapts to any part constraint for grasping","Unified framework for open-vocab part grasping with dual arms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the VLM and the stacked segmentation models, object first then part, return the user-intended functional part as a geometrically coherent point-cloud mask; when that mask is wrong, fragmented, or off-target, the part-guided diffusion is handed an invalid constraint and cannot recover, and the paper does not isolate this upstream failure rate.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot dual-arm grasping via part-guided diffusion","Part labels guide diffusion for 87.6% grasp success","VLM+diffusion: open-vocab grasps with dual-arm coordination","Diffusion adapts to any part constraint for grasping","Unified framework for open-vocab part grasping with dual arms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000417,"raw_usage":{"total_tokens":2180,"prompt_tokens":1004,"completion_tokens":1176,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1090}},"tokens_in":620,"tokens_out":1176,"duration_ms":8547,"temperature":1.0,"reasoning_tokens":1090,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:30:57.810892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical objects and instructions with ground-truth part masks substituted for the VLM-segmentation output; if success rates stay near the reported levels, the VLM-driven constraint stage is not doing the work, whereas a large drop would confirm that mask quality is the bottleneck. A stage-wise failure log that counts missed or malformed part masks separately from grasp-pose collisions would also settle where the pipeline breaks.","supporting_citations":[{"cited_title":"Constrained 6-dof grasp generation on complex shapes for improved dual-arm manipulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the Part-Guided Diffusion strategy inside a Constrained Grasp Diffusion Field, the core grasp-generation engine that consumes VLM-derived part constraints."},{"cited_title":"Going denser with open-vocabulary part segmentation,","cited_arxiv_id":null,"evidence_quote":"VLPart performs open-vocabulary part segmentation on the object mask, producing the target part point cloud that conditions diffusion."}],"review_version":1}