{"id":"afd32cd9-6018-484a-add6-21c72db5eb56","arxiv_id":"2608.10383","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A real-world pipeline trains a diffusion model on teleoperated bimanual demonstrations and adds force-guided refinement to grasp large objects from a single RGB-D view.","lead":"This paper presents a system that lets a two-armed humanoid robot grasp large objects with dexterous hands, planning the grasp from a single camera view and refining it with force and touch sensors during execution. On 7 unseen objects it reports 61.9% average lift-and-hold success, beating a simulation-based bimanual baseline (28.6%) and an RL-based baseline (37.1%) in real-world trials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Force-guided refinement may be the actual grasping policy; the paper never reports success of the raw predicted pose, so the 61.9% cannot be attributed to the learned diffusion model.","rationale":"The reader's weakest assumption concerned the recovery range of force-guided refinement; I partially agree and sharpen it into a missing-condition problem. Section III.C describes a two-stage reactive controller that can grasp objects even with a mediocre initial pose, because it closes arms until physically blocked and bends fingers until contact. The paper's Table III ablations isolate the removal of arm squeezing and finger adjustment, but never the removal of the entire refinement stage. Without this condition, the 61.9% success rate conflates the learned model's contribution with the handcrafted controller's. This matters because the central claim is about single-view grasp synthesis: if the raw DDPM output is not executable, the system's success is due to reactive clamping, not to the learned synthesis, and the comparison against baselines—which do not include such refinement—is not a fair test of grasp generation. The paper does not report error bars or artifact verification, so the raw-pose condition is a cheap and decisive experiment. If the authors run the released model with refinement disabled and find success remains substantial, the recovery-range assumption holds; if not, the headline quantitative claim must be weakened to 'diffusion predicts an approximate palm placement, refined by force control.' Since the evidence is currently absent, the conditional verdict is appropriate.","tokens_in":11821,"tokens_out":9756,"duration_ms":107968,"concrete_test":"Run the released model and code on the same 7 test objects under three conditions: (A) full pipeline; (B) raw predicted joint angles executed without any force-guided refinement (no arm squeeze, no finger bending); and (C) handcrafted centroid initialization plus force-guided refinement (the paper's 'w/o DDPM' ablation). If condition B's success is near zero while C approaches the 61.9% of A, the diffusion model contributes little and the refinement is the grasping policy. Also log pre- and post-refinement palm positions on successful A trials; if median displacement exceeds ~5 cm, the raw predictions are outside the recovery basin and the headline claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that a DDPM trained on teleoperated demonstrations synthesizes executable bimanual grasps from a single-view point cloud, with 61.9% real-world success. But the execution pipeline in Section III.C adds 'move arms inward until blocked' and 'bend fingers until contact' — a reactive policy that can convert a rough palm placement into a clamped grasp regardless of the predicted finger configuration. The ablations in Table III show that removing arm squeezing or finger adjustment drops success from 61.9% to 42.86% and 34.29%, respectively, but the paper never reports the success of executing the raw predicted pose with zero force-guided refinement. That missing condition is load-bearing: if the raw predicted pose is rarely executable on its own, then the learned model is not actually synthesizing grasps; the handcrafted squeezing policy is doing the grasping, and the baselines (which lack this reactive refinement) are compared against a system whose success is due to a different mechanism. Figure 7(d) shows that when the predicted placement is wrong (grasping one end of a heavy box), refinement cannot recover, confirming that the recovery range is finite. Without the no-refinement condition, the quantitative contribution of the diffusion model is unidentified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a three-phase system for bimanual dexterous grasping of large objects on a real dual-arm humanoid. A teleoperation dataset (353 grasps on 33 objects) is collected; a DDPM is trained to predict a 26-dimensional arm-and-hand configuration from a SAM-segmented single-view point cloud; and execution combines motion planning with a force-guided refinement stage that moves arms inward until blocked and bends fingers until contact. The authors report a 61.90% mean lift-and-hold success rate over 7 unseen objects (15 trials each), with ablations degrading when components are removed, and they release code and a dataset.","tokens_in":12144,"tokens_out":9440,"duration_ms":105312,"significance":"The physical real-robot evaluation on unseen objects is a genuine strength: success is defined by lift-and-hold, the per-object rates and totals are internally consistent, and the ablations generally degrade monotonically as modules are removed. If the attribution to the learned model were cleanly identified, the system would be a useful step toward real-world bimanual grasping without full object meshes. However, because the force-guided refinement is a handcrafted reactive policy and is never removed in a no-refinement condition, the reported numbers currently cannot be attributed to the diffusion model; the lack of uncertainty quantification also weakens the comparative claims. These issues are fixable with additional experiments rather than being fundamental.","major_comments":[{"comment":"The central attribution of the 61.90% success rate to the learned diffusion model is unidentified because no condition executes the raw predicted pose h* without force-guided refinement. Section III.C adds 'move arms inward until blocked' and 'bend fingers until contact' (Eq. (9)), a reactive policy that can convert a coarse palm placement into a clamped grasp regardless of the predicted finger configuration. Table III ablates arm squeezing and finger adjustment separately, but never removes both; the 'ours w/o motion planning' row still includes the refinement stage. The manuscript also does not report the success of the handcrafted centroid-positioning baseline when followed by the same force-guided refinement. Figure 4 shows generated poses 'without force-guided refinement' but gives no quantitative success for them. Without these conditions, the quantitative contribution of the diffusion model over the reactive policy is unidentified, and the comparison with baselines that lack the refinement stage is confounded. Please add: (i) raw DDPM output executed open-loop with no force-guided refinement, and (ii) the handcrafted baseline followed by the same refinement.","section":"III.C and Table III"},{"comment":"The experimental comparison lacks uncertainty quantification. Each per-object success rate is an average of 15 binary trials, so with only 7 objects many per-object gaps are within sampling noise; for example, 73.33% versus 40.00% on the basin corresponds to 11 versus 6 successes out of 15. The aggregate differences may be real, but the paper needs raw trial counts, confidence intervals, or statistical tests (for example, Fisher's exact test per object or a paired test over object means) to support the claim that the proposed method consistently outperforms the baselines.","section":"IV.B and Table II"},{"comment":"The ablation 'ours w/o DDPM' replaces the learned predictor with a handcrafted symmetric-centroid positioning baseline, not with a non-diffusion learned model. This shows that some learned prediction is useful, but it does not isolate the diffusion formulation. Since the paper's stated contribution is a DDPM-based generator, a comparison with a non-diffusion learned baseline trained on the same data (for example, an MLP regressor or a VAE) is needed to support the claim that the diffusion model per se is responsible for the improvement.","section":"IV.C and Table III"}],"minor_comments":[{"comment":"The table columns 'box 1', 'box 2', and 'box 3' are not self-explanatory; the text maps only box 1 and box 2 to object names, so please add a legend or use the physical object names consistently in the tables.","section":"Tables II and III"},{"comment":"Equation (8) defines two contact thresholds, tau_th and delta_th, but Table I reports only a single force threshold (2.9 N). Please specify the tactile threshold delta_th and give units or normalization for the tactile readings T_{i,j}.","section":"III.C and Table I"},{"comment":"The baseline descriptions do not state whether GraspNet and DexGraspAnything outputs are also given a pre-grasp offset or any collision-aware motion planning; please clarify whether the reported baseline failures include motion-level failures, to rule out trivial execution artifacts.","section":"IV.B"},{"comment":"The title and abstract emphasize 'single-view observations,' and the paper explicitly acknowledges in Section III.C that force feedback is used during execution. Please make the split of responsibility clearer in the abstract: the learned synthesis is single-view, while the full system is a feedback policy. This would prevent a reader from inferring that the demonstrated success comes from vision alone.","section":"III.C"}],"recommendation":"major_revision","confidential_remarks":"The missing raw-prediction condition is the primary blocker. I would ask the authors to add an ablation that executes the raw DDPM output with neither arm squeezing nor finger adjustment, and to report the handcrafted baseline with the same force-guided refinement. In the same revision, I would require per-trial counts and confidence intervals or significance tests for the headline comparison. If the missing condition cannot be run, the attribution claims should be softened and the system presented as a learned-prediction-plus-reactive-refinement policy rather than a pure learned synthesis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a genuine real-world bimanual grasping system with a new teleoperated dataset and a plausible pipeline, and the 61.9% success over 7 unseen objects is worth taking seriously. But the stress-test concern is right: the authors never report success of the raw DDPM output executed without force-guided refinement, so the headline attribution of the result to the learned model is not supported by the data. The system works; we don't know how much of the work is the diffusion model versus the reactive squeezing and finger-bending.\n\nWhat's actually new: single-view segmented point cloud to 26-DoF bimanual joint configuration via DDPM, executed on a real dual-arm humanoid with dexterous hands, plus a 353-grasp multimodal dataset (joint angles, RGB-D, force/tactile). That integration hasn't been shown before on real hardware. The real-robot evaluation with lift-and-hold criteria, object-level split, and multiple poses per object is solid work. Tables are internally consistent, and the failure analysis (Figure 7) is honest.\n\nSoft spots, in proportion. First and most important: the missing no-refinement condition. The ablations show that removing arm squeezing or finger adjustment drops success from 61.9% to 42.9% and 34.3%, but both those variants still retain most of the reactive refinement. Without a condition that executes the predicted pose with plain joint control, you can't know whether the DDPM is finding a good grasp or just a rough palm placement that the handcrafted policy then clamps into a hold. The paper's own Figure 7(d) shows the refinement can't recover when the predicted placement is wrong, which confirms the recovery range is finite and makes the missing condition more urgent. Second, the generator ablation baseline 'without DDPM' is a handcrafted symmetric open-palm posture, which is a very low bar; it doesn't isolate the learned model's value. Third, no confidence intervals, no significance tests, n=15 per object, only 7 unseen objects, all under fixed lighting and uniform background. Those would be needed to claim 61.9% beats 28.6% etc. with any precision. The ViSiL-HD baseline adaptation (no support from underneath, tabletop placement) is acknowledged but makes that comparison less clean. Calling the waypoint interpolation 'motion planning' is an overstatement. The code/data link is unverified, but the experiments are physical, so that's a minor point.\n\nBottom line: this deserves a serious referee. The central engineering result is believable and useful for people working on bimanual dexterous manipulation. But the learned-generator claim needs a no-refinement condition, uncertainty quantification, and a stronger ablation baseline before the numbers can be taken at face value.","headline":"A real bimanual grasping system with honest real-robot experiments, but the missing no-refinement condition leaves the DDPM's contribution unidentified.","tokens_in":12739,"tokens_out":3003,"would_cite":true,"duration_ms":31706,"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":"A diffusion model plus force-guided refinement reaches 61.9% success grasping large objects bimanually from a single RGB-D view.","keywords":["bimanual grasping","dexterous hands","diffusion models","single-view point cloud","force-guided refinement","teleoperation dataset","real-world robotic manipulation"],"falsifier":"Run the pipeline on a set of objects whose center of mass is deliberately offset from the geometric center—for example, a long box with a heavy weight fixed at one end. If the single-view diffusion prediction places the hands near one end and the monotonic refinement cannot shift them toward the center of mass, the lift-and-hold success rate should drop markedly below the reported 61.9%; such a result would show that single-view sufficiency does not hold for uneven mass distributions.","tokens_in":11574,"feed_emoji":"👐","tokens_out":9447,"duration_ms":81883,"temperature":0.7,"pith_summary":"The paper claims that a humanoid robot with two dexterous hands can grasp large, everyday objects—bins, boxes, buckets, balls—from a single RGB-D image, without needing a full 3D model of the object. The method trains a diffusion model on teleoperated demonstrations to predict arm and hand joint angles from a segmented partial point cloud, then executes the predicted grasp with motion planning and a force-guided closing routine. In real-world trials on seven unseen objects, the pipeline lifts and holds objects in 61.9% of attempts, outperforming the compared bimanual and single-hand baselines. The authors conclude that partial-view perception plus reactive contact refinement is sufficient for stable cooperative bimanual grasping of large objects.","feed_headline":"One RGB-D view drives bimanual dexterous grasps at 61.9%","feed_subtitle":"A diffusion model plus force-guided refinement lifts unseen boxes, bins, and balls without full 3D models.","key_machinery":"The load-bearing components are three. First, a teleoperated dataset of 353 grasp sequences on 33 large objects, collected from a humanoid robot with a head-mounted RGB-D camera and teleoperation, recording joint angles, RGB-D images, joint torques, and tactile signals, provides the training signal. Second, a DDPM with a 4-layer MLP denoiser, conditioned on a PointNet++ object feature, generates a 26-dimensional joint configuration from Gaussian noise over 200 denoising steps, with joint limit clipping for hardware feasibility. Third, the execution strategy computes a waypoint-based trajectory that lifts the arms laterally and approaches along the palm normal, then applies force-guided refinement: each palm moves toward the object midpoint until blocked, and each finger bends incrementally using torque and tactile thresholds until contact. The refinement is what converts a predicted pose into a stable grasp by compensating for geometric deviations not visible in a single view.","core_discovery":"The central claim is that a Denoising Diffusion Probabilistic Model, conditioned on a 1024-dimensional PointNet++ feature of a single-view segmented point cloud, can synthesize a 26-dimensional bimanual grasp configuration (14 arm joints and 12 hand joints), and that executing this configuration with a two-stage force-guided refinement—moving arms inward until blocked and bending fingers until contact—yields physically stable grasps on objects the robot has never seen. The paper reports a mean lift-and-hold success rate of 61.9% across seven test objects, with each object tested in five poses and three attempts per pose, exceeding BimanGrasp-DDPM (28.6%), ViSiL-HD (37.1%), GraspNet (21.9%), and DexGraspAnything (15.7%). Ablations show that removing the arm-squeezing stage, the finger-adjustment stage, or the motion planning stage each substantially reduces success, and replacing the diffusion model with a handcrafted symmetric-open-palm baseline also degrades performance.","pith_inferences":["A testable extension of the paper's claim is that learning the contact-refinement policy with tactile feedback—rather than the handcrafted monotonic closing—would raise success rates on off-center or heavy objects; the paper's own failure case (d) points to this limit.","If the central claim holds, the design priority for bimanual dexterous grasping shifts from full-shape reconstruction to the reactive contact loop; the single-view point cloud need only approximate the grasp region.","Because the dataset covers 33 objects but the object-level split leaves only seven for testing, measuring success across a larger suite of objects would clarify how quickly the diffusion model's grasp synthesis degrades with novel geometry.","The handcrafted waypoint planner could be replaced by a learned motion policy that reasons about free space, potentially reducing failures from table collisions and torso obstruction that the paper documents."],"forward_implications":["Bimanual grasp synthesis for large objects no longer requires a complete 3D mesh; a single head-mounted RGB-D frame provides enough information to generate an executable grasp configuration.","The contact refinement stage is essential to real-world stability: removing arm squeezing or finger adjustment drops success from 61.9% to 42.9% or 34.3%, respectively.","A teleoperated, real-world dataset can support learning of physically reachable bimanual grasp poses, addressing a gap left by simulation-only methods.","The framework leaves room for extension to more complex scenes via open-vocabulary detection and collision-aware motion planning, as the authors note."],"supporting_citations":[{"why":"This is the simulation-only bimanual grasp synthesis baseline that the paper extends and that motivates the need for real-world validation.","marker":"[7]"},{"why":"This supplies the object mask that converts the RGB image into the partial point cloud used for grasp generation.","marker":"[39]"},{"why":"This encodes the partial point cloud into the 1024-dimensional object feature that conditions the diffusion model.","marker":"[40]"},{"why":"This provides the generative diffusion framework that maps noise to a joint configuration.","marker":"[41]"},{"why":"This is a single-hand baseline that, together with the other baselines, demonstrates the benefit of bimanual grasping for large objects.","marker":"[42]"},{"why":"This is a single dexterous-hand baseline used to show that large, simple objects need two hands.","marker":"[20]"},{"why":"This is a real-world dual-arm dexterous manipulation baseline whose limited object generality motivates the proposed method.","marker":"[38]"}],"fun_headline_variants":["One RGB-D view: bimanual grasp hits 61.9% in real world","Diffusion model turns single view into 61.9% bimanual grasps","Bimanual grasp from one view: 61.9% success on unseen objects","Real-world bimanual grasp: 61.9% with diffusion model","Single-view bimanual grasp achieves 61.9% on novel objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The diffusion model's predicted joint configuration must lie within the recovery range of the handcrafted force-guided refinement, meaning the predicted hand and arm positions must be close enough to a stable grasp that moving the arms inward and bending the fingers monotonically can close the remaining gap; if the prediction is far off, as when a heavy object is grasped at one end, refinement cannot prevent slippage.","fun_headline_variants_meta":{"raw":{"variants":["One RGB-D view: bimanual grasp hits 61.9% in real world","Diffusion model turns single view into 61.9% bimanual grasps","Bimanual grasp from one view: 61.9% success on unseen objects","Real-world bimanual grasp: 61.9% with diffusion model","Single-view bimanual grasp achieves 61.9% on novel objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3041,"prompt_tokens":952,"completion_tokens":2089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1981}},"tokens_in":568,"tokens_out":2089,"duration_ms":16331,"temperature":1.0,"reasoning_tokens":1981,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:50:22.865001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a set of objects whose center of mass is deliberately offset from the geometric center—for example, a long box with a heavy weight fixed at one end. If the single-view diffusion prediction places the hands near one end and the monotonic refinement cannot shift them toward the center of mass, the lift-and-hold success rate should drop markedly below the reported 61.9%; such a result would show that single-view sufficiency does not hold for uneven mass distributions.","supporting_citations":[{"cited_title":"Bimanual grasp synthesis for dexterous robot hands,","cited_arxiv_id":null,"evidence_quote":"This is the simulation-only bimanual grasp synthesis baseline that the paper extends and that motivates the need for real-world validation."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"This supplies the object mask that converts the RGB image into the partial point cloud used for grasp generation."},{"cited_title":"Pointnet++: Deep hierar- chical feature learning on point sets in a metric space,","cited_arxiv_id":null,"evidence_quote":"This encodes the partial point cloud into the 1024-dimensional object feature that conditions the diffusion model."},{"cited_title":"Graspnet-1billion: A large- scale benchmark for general object grasping,","cited_arxiv_id":null,"evidence_quote":"This is a single-hand baseline that, together with the other baselines, demonstrates the benefit of bimanual grasping for large objects."},{"cited_title":"Dexgrasp anything: Towards universal robotic dexterous grasping with physics awareness,","cited_arxiv_id":null,"evidence_quote":"This is a single dexterous-hand baseline used to show that large, simple objects need two hands."}],"review_version":1}