{"id":"cc23e21b-eca8-4e79-aec1-6d1cc648375f","arxiv_id":"2411.15903","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"BimanGrasp produces a large-scale simulated dataset of bimanual dexterous grasps and a diffusion model that synthesizes them at quasi-real-time speeds.","lead":"This paper presents BimanGrasp, a method that synthesizes two-handed grasps for dexterous robot hands on 3D objects, verified in physics simulation. It also contributes a dataset of over 150,000 validated bimanual grasps and a diffusion model that generates new grasps much faster than the optimizer.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported DDPM success is always after a 100-step energy-based post-processing refinement, and no ablation isolates the diffusion model from the optimizer; without such an ablation the central acceleration/comparability claim is not yet supported.","rationale":"I read the paper as making two coupled claims: the BimanGrasp optimizer plus Isaac Gym labels a useful dataset, and the DDPM trained on that dataset matches optimizer quality at lower cost. The simulator-transfer limitation is real but is a standard future-work item for grasp datasets and is acknowledged in Sec. IV-C. The more actionable gap is internal: every DDPM success number in the paper is produced by a pipeline that finishes with an energy-based optimization using the same object mesh, penetration terms, and quality objective as the optimizer. With the DDPM result at rho=2500 landing at 54.06% versus 54.03% for the optimizer, the numbers are consistent with the post-processing dominating the outcome. The paper would be substantially stronger, and the central claim properly attributed, if the authors reported (i) raw DDPM sample success before post-processing and (ii) the same 100-step refinement from non-learned initializations. This is a concrete, cheap experiment. I therefore keep the reader's CONDITIONAL verdict: condition acceptance on this ablation, and on correcting the density typo.","tokens_in":49,"tokens_out":6850,"duration_ms":128558,"concrete_test":"Using the paper's held-out protocol (225 unseen GSO objects, 500 grasps per object, Isaac Gym), compare: (a) raw BimanGrasp-DDPM samples without post-processing; (b) DDPM samples followed by the paper's 100-step refinement; (c) 100-step refinement from random initial grasps sampled as in Sec. III-B, with no learned model; and (d) 100-step refinement from a known weak baseline such as CVAE or Uni2Bim. Evaluate all four at rho=5000, 2500, and 500 kg/m^3, correcting the density labeling in Sec. IV-B. If (c) matches (b), the DDPM contributes no grasp-quality improvement; if (a) is far below (b), the reported success is attributable to post-processing rather than to the diffusion model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central DDPM claim (Sec. IV-B, abstract) is evaluated on outputs that have been post-processed by optimizing the same energy terms used by the BimanGrasp optimizer for 100 steps (Sec. III-D). The reported DDPM success rates (42.39%, 54.06%, 69.87%) are nearly equal to the BimanGrasp optimizer rates at the corresponding densities (41.02%, 54.03%, 71.42% in Table II), and the density label for the 69.87% row is internally inconsistent (the text repeats rho=2500, suggesting a typo for rho=500). Because no baseline runs the same 100-step refinement from random or baseline-generated initial grasps, the paper does not establish that the diffusion model is what makes generation fast or successful; the improvement could come entirely from the energy-based post-processing. This is more directly load-bearing than the simulator-transfer question: even granting Isaac Gym as ground truth, the attribution of the central result to the DDPM is unproven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for synthesizing stable bimanual grasps for two Shadow Hands. First, it introduces BimanGrasp, a stochastic-optimization algorithm that minimizes a weighted sum of energy terms capturing hand-object distance, force closure, wrench volume, penetration, and joint limits. Second, it generates a large dataset of grasp poses for 900 GSO objects, labels them with success/failure in Isaac Gym, and releases the successful subset as the BimanGrasp-Dataset. Third, it trains a conditional diffusion model (BimanGrasp-DDPM) on this dataset, with a short 100-step energy-based post-processing refinement, and reports success rates comparable to the optimizer at lower computational cost. The paper also compares bimanual versus unimanual grasping and reports robustness to object density and friction coefficient.","tokens_in":12678,"tokens_out":4320,"duration_ms":39901,"significance":"If the central claims hold, this is a useful contribution to dexterous manipulation: it provides a first large-scale bimanual dexterous-grasp dataset, demonstrates a joint-optimization formulation for two high-DoF hands, and shows that a diffusion model can propose grasps for unseen objects. Strengths include the physics verification protocol in Isaac Gym with randomized gravity directions, the dataset scale, and the clear presentation of the algorithmic pipeline. However, the headline claim about BimanGrasp-DDPM being comparable to the optimizer at much lower cost is currently confounded by the energy-based post-processing step, and several reproducibility-critical details are missing. The paper deserves revision rather than rejection because the issues are addressable with additional experiments and reporting.","major_comments":[{"comment":"The energy function is defined as a weighted sum of the terms in Table I, but the weights are never reported. Without these weights, the BimanGrasp optimization is not reproducible, and one cannot assess how the 100-step DDPM post-processing in Sec. III-D relates to the full 10000-step optimizer objective. Please report the exact weights for all seven terms, along with the penetration threshold δ, the distance threshold ϵ, and the selection rule for the eight contact points.","section":"Sec. III-B, Table I"},{"comment":"The core claim that BimanGrasp-DDPM generates grasps with success comparable to the optimizer is not yet supported, because all reported DDPM numbers are obtained after 100 steps of energy-based post-processing using the same energy terms as the BimanGrasp optimizer. The DDPM success rates (42.39%, 54.06%, 69.87%) closely match the optimizer rates at the corresponding densities (41.02%, 54.03%, 71.42% in Table II), which is consistent with the post-processing, not the diffusion model, carrying the performance. Please add an ablation: report raw DDPM success before post-processing, success of 100-step energy refinement from random initial grasps, and success of 100-step refinement from baseline-generated grasps. Without these conditions, the acceleration and comparability claims cannot be attributed to the DDPM.","section":"Sec. III-D and Sec. IV-B"},{"comment":"The density label for the headline success rate is internally inconsistent: the text reports 69.87% for ρ = 2500 kg·m⁻³, but 54.06% was already reported for that density, and the 69.87% value is instead consistent with the ρ = 500 row of Table II. This typo affects the abstract's central number and must be corrected, and the reported DDPM rates should be rechecked against the experimental records.","section":"Sec. IV-B"},{"comment":"The two learned baselines, CVAE and Uni2Bim(dm), are not described in enough detail to judge fairness. The manuscript does not specify the CVAE architecture, training procedure, conditioning input, or whether either baseline receives the same 100-step energy-based post-processing as BimanGrasp-DDPM. If the baselines are evaluated without post-processing, the comparison conflates model quality with the refinement step. Please provide full implementation details and, ideally, also evaluate the baselines with the same post-processing protocol.","section":"Sec. IV-B"},{"comment":"All success rates are reported as single numbers without error bars, confidence intervals, or object-level variance. Given that the evaluation averages over 900 objects and 500 grasps per object, object-to-object variance is likely substantial and could change the conclusions about 'consistently higher' and 'comparable' performance. Please report per-object mean and standard deviation (or confidence intervals) for the main comparisons.","section":"Tables II and III, Sec. IV-B"}],"minor_comments":[{"comment":"The abstract states the dataset contains 'over 150k verified grasps,' while Sec. IV-A reports synthesizing 450k bimanual grasps (900 objects × 500 poses). Please clarify whether the dataset contains 150k verified grasps and what happened to the remaining synthesized grasps.","section":"Abstract and Sec. IV-A"},{"comment":"The reference list in the sentence on variational autoencoders reads '[23], [23]–[25]', which appears to be a duplicate citation typo and should be cleaned up.","section":"Sec. II-A"},{"comment":"The definition of d(p, O) is written as min_{q∈O}(p, q); this should be min_{q∈O} ||p − q||, and the sentence about the penetration terms 'unless it is lower than a fixed small threshold ϵ' is ambiguous about which quantity is thresholded.","section":"Table I"},{"comment":"The friction coefficient is fixed at 3, which is high for typical objects and hand surfaces. Since Table III later shows sensitivity to friction, please justify this choice or add a brief discussion of its effect on the dataset labels and on transfer to real hardware.","section":"Sec. III-C"},{"comment":"The computational cost paragraph reports 170 GB GPU memory and 117 minutes per 4,500 grasps for dataset generation, and 8.19 seconds for 64 parallel DDPM inferences, but it does not give the per-grasp time for the full DDPM-plus-post-processing pipeline or for the BimanGrasp optimizer. Please include a direct speed comparison to support the 'significant acceleration' claim.","section":"Sec. IV-C"},{"comment":"The text refers to 'Bimanual-DDPM' in one place; the model name elsewhere is BimanGrasp-DDPM. Also, Fig. 7 would benefit from error bars to support the claim that bimanual grasps outperform unimanual baselines across all diameter bins.","section":"Sec. IV-B and Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of RA-L and the dataset could be a useful community resource. My main concern is not simulator transfer but attribution: the DDPM result is evaluated after the same energy-based refinement used by the optimizer, so the paper's central speed/quality claim requires an ablation that isolates the diffusion model. The missing energy weights are also a reproducibility blocker. These are fixable with additional experiments and reporting, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core deliverable here is BimanGrasp-Dataset: about 150k simulated bimanual Shadow Hand grasps on 900 GSO objects, each verified in Isaac Gym by lifting and holding with randomized gravity directions. The optimization pipeline that produced it is a reasonable extension of DexGraspNet to two hands, with force-closure, wrench-volume, penetration, and joint-limit terms. The comparison against single-hand baselines shows a plausible advantage for large objects. That part is a solid resource for the bimanual manipulation community.\n\nThe diffusion model story is where the paper gets shaky. Every reported success rate (42.39%, 54.06%, 69.87%) comes from DDPM outputs that were then refined by 100 steps of the very same energy function that BimanGrasp optimizes (Sec. III-D). The raw DDPM output is never reported. The reported rates are close to the BimanGrasp optimizer's rates in Table II, which is suspicious. The stress-test note is right: without a baseline that applies the same 100-step refinement to random or baseline-generated initial grasps, you cannot attribute the success or the speed to the diffusion model. The speed comparison is also muddled; 8.19 seconds for 64 grasps on an RTX 4090 is not obviously \"quasi-real time,\" and the optimizer's timing is not given on the same hardware.\n\nOther soft spots, in descending order of importance: the energy weights in Table I are never reported, so the optimizer is not reproducible as described. There are no error bars anywhere, and the density row for 69.87% repeats rho=2500, likely a typo for 500. The \"first\" claim is overstated because ArtiGrasp [7] does bimanual dexterous grasping, though its focus is articulation. And everything is simulated; the authors explicitly defer real-robot validation.\n\nOn balance, I would send this to peer review if the venue cares about datasets and simulation resources. The optimizer and dataset will be useful even if the DDPM claim is later qualified. But the referee must push for the ablation and the missing details. I would not take the DDPM numbers at face value.","headline":"Useful bimanual dexterous grasp optimizer and dataset, but the DDPM evaluation is confounded by energy-based post-processing and needs an ablation.","tokens_in":13197,"tokens_out":2448,"would_cite":true,"duration_ms":21767,"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":"Bimanual dexterous grasp synthesis can be solved by optimizing an energy function and then accelerated with a diffusion model trained on verified grasps.","keywords":["bimanual grasping","dexterous manipulation","grasp synthesis","diffusion model","grasp dataset","force closure","physics simulation","robot hands"],"falsifier":"Repeat the evaluation on a real bimanual humanoid with two 22-DoF hands across the same object set; if hardware success rates fall well below the simulated rates, the simulator's contact model is not transferable and the central claim fails in practice. Even inside simulation, re-running the verification with a different contact friction model or solver and checking whether the bimanual-versus-unimanual ranking persists would test the robustness of the comparison.","tokens_in":12280,"feed_emoji":"🤖","tokens_out":8579,"duration_ms":71805,"temperature":0.7,"pith_summary":"Most robotic grasp synthesis assumes one hand, leaving large and heavy objects that naturally require two hands largely unhandled. This paper tries to close that gap by proposing an optimization-based pipeline that searches the combined configuration space of two dexterous hands for stable bimanual grasps on arbitrary 3D object meshes. Each candidate grasp is checked in a physics simulator, producing a dataset of over 150,000 verified bimanual grasps across 900 objects. The paper then trains a denoising diffusion model on those verified grasps, and reports that it generates grasps for unseen objects at a success rate comparable to the optimizer's while being far faster. If this holds, it gives humanoid robots a general, fast bimanual grasping skill for everyday objects.","feed_headline":"Diffusion model generates two-hand grasps at 69.87% success","feed_subtitle":"A 150k-grasp dataset lets a diffusion model match an optimizer's two-hand grasp success at a fraction of the compute.","key_machinery":"The load-bearing object is the energy function in Table I, whose terms are $E_{\\mathrm{dis}}$ (distance between hand surface points and object), $E_{\\mathrm{fc}}$ (force closure via the norm of the grasp matrix $G$ built from 8 contact points), $E_{\\mathrm{vew}}$ (wrench-ellipse volume, keeping $GG^T$ well conditioned), three penetration penalties $E_{\\mathrm{objpen}}$, $E_{\\mathrm{selfpen}}$, $E_{\\mathrm{bimpen}}$, and $E_{\\mathrm{joint}}$ for joint-limit violations. Minimizing this function over the 56-dimensional action space of two hands, each with 22 joint angles and a rigid-body pose, defines the BimanGrasp search. The same energy function reappears as a short post-processing refinement after the diffusion model generates a candidate, which is how penetration errors from the generative step are cleaned up. The generative model itself is a denoising diffusion probabilistic model conditioned on point-cloud features of the object.","core_discovery":"The central claim is that bimanual grasp synthesis for dexterous hands can be made reliable by optimizing a hand-crafted energy function, and then made fast by learning from the optimized results. The energy function rewards closeness to the object surface, force closure measured through an 8-contact grasp matrix $G$, robustness of the wrench ellipse, and penalties for hand-object, self-, and inter-hand penetration plus joint-limit violations; minimizing it with a stochastic optimizer yields the BimanGrasp algorithm. Physical verification in simulation labels which of the produced grasps can lift and hold an object for two seconds under randomized gravity directions. The verified grasps form a dataset that trains BimanGrasp-DDPM, a conditional diffusion model that turns object point-cloud features into new bimanual grasp poses; a short post-processing optimization removes penetrations. The reported result is that the diffusion model reaches a 69.87% verification success rate, close to the optimizer's rate, while generating 64 grasps in parallel in 8.19 seconds on a single commercial GPU.","pith_inferences":["If the simulator's contact model transfers to hardware, the remaining bottleneck is not the generative model but the post-processing step: even the fast model still spends optimization steps per grasp to remove penetrations, so a diffusion model with built-in physical constraints could make generation truly single-shot.","The same dataset could be reused beyond grasp synthesis, for example as supervision for bimanual manipulation policies or as a prior for tasks requiring coordinated finger placement; the paper does not train such policies, so this is an extension the authors leave implicit.","A sim-to-real gap is the untested risk: the 69.87% success rate is measured in the same simulator used to label the training data, so hardware deployment would require a separate transfer evaluation."],"forward_implications":["If correct, robot manipulators gain a principled route from object mesh to coordinated two-hand grasp, including objects as large as 0.7 m in diameter that unimanual methods almost never grasp.","Jointly optimizing the two hands matters: the same pipeline run as two independent single-hand optimizers succeeds less often at every tested object density.","A learned generator can replace most of the expensive optimization: BimanGrasp-DDPM matches the optimizer's success rate at a fraction of the compute and works on unseen objects.","The verified-grasp dataset is reusable: training on only 75% of the objects still transfers to the remaining 25% and to objects from other benchmark datasets.","The method's advantage grows with object mass: bimanual success degrades more gracefully than unimanual as density increases."],"supporting_citations":[{"why":"Supplies the differentiable force closure estimator and stochastic optimization approach that the BimanGrasp energy terms build on.","marker":"[12]"},{"why":"Provides the unimanual grasp synthesis baseline and the initialization procedure that BimanGrasp adapts to two hands.","marker":"[4]"},{"why":"Provides the physics simulation environment used to verify grasp success and label the dataset.","marker":"[13]"},{"why":"Supplies the scanned household object meshes used for training and evaluation.","marker":"[40]"},{"why":"Defines the denoising diffusion probabilistic model architecture used by BimanGrasp-DDPM.","marker":"[41]"},{"why":"Supplies the point-cloud encoder that produces the object-conditioning feature for the diffusion model.","marker":"[42]"},{"why":"Provides the unimanual generative grasping approach adapted as the Uni2Bim(dm) baseline.","marker":"[26]"},{"why":"Provides the conditional variational autoencoder architecture used as the CVAE baseline.","marker":"[23]"}],"fun_headline_variants":["Diffusion model matches optimizer for two-hand grasps at 69.87% success","150k verified two-hand grasps unlock fast diffusion-based synthesis","Diffusion model cuts two-hand grasp compute while keeping 69.87% accuracy","First large-scale bimanual grasp dataset powers fast diffusion synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The physics simulator used for verification models real contact and friction accurately enough that a grasp labeled stable in simulation will also hold on physical hardware; the paper does not test on a real robot.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model matches optimizer for two-hand grasps at 69.87% success","150k verified two-hand grasps unlock fast diffusion-based synthesis","Diffusion model cuts two-hand grasp compute while keeping 69.87% accuracy","First large-scale bimanual grasp dataset powers fast diffusion synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000676,"raw_usage":{"total_tokens":3089,"prompt_tokens":973,"completion_tokens":2116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2036}},"tokens_in":589,"tokens_out":2116,"duration_ms":13452,"temperature":1.0,"reasoning_tokens":2036,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:44:52.073507+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the evaluation on a real bimanual humanoid with two 22-DoF hands across the same object set; if hardware success rates fall well below the simulated rates, the simulator's contact model is not transferable and the central claim fails in practice. Even inside simulation, re-running the verification with a different contact friction model or solver and checking whether the bimanual-versus-unimanual ranking persists would test the robustness of the comparison.","supporting_citations":[{"cited_title":"Synthesizing diverse and physically stable grasps with arbitrary hand structures using differ- entiable force closure estimator,","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable force closure estimator and stochastic optimization approach that the BimanGrasp energy terms build on."},{"cited_title":"Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation,","cited_arxiv_id":null,"evidence_quote":"Provides the unimanual grasp synthesis baseline and the initialization procedure that BimanGrasp adapts to two hands."},{"cited_title":"Google scanned objects: A high- quality dataset of 3d scanned household items,","cited_arxiv_id":null,"evidence_quote":"Supplies the scanned household object meshes used for training and evaluation."},{"cited_title":"UGG: Unified Generative Grasping","cited_arxiv_id":"2311.16917","evidence_quote":"Provides the unimanual generative grasping approach adapted as the Uni2Bim(dm) baseline."},{"cited_title":"Hand-object contact con- sistency reasoning for human grasps generation,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional variational autoencoder architecture used as the CVAE baseline."}],"review_version":1}