{"id":"c6acb0eb-481f-4d5f-8447-88ea7db7c4c7","arxiv_id":"2411.10049","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SPLIT predicts SE(3) pose scores from local geometry in the sample frame, enabling a single diffusion model to generate grasps, handle poses, and upright placement poses for mug manipulation.","lead":"This paper presents SPLIT, a diffusion model that generates robot poses (grasps, handle positions, placement orientations) directly from a 3D scene by reading only the local geometry around each candidate pose. The same model can solve several manipulation subtasks at once, such as picking up and hanging a mug, which could allow robots to replace many task-specific perception modules with one learned system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The locality assumption is load-bearing, but the architecture can access global geometry via multi-scale feature grids, so the experiments do not isolate local context; the hidden-handle failure mode confirms non-local priors are used.","rationale":"The reader correctly identifies Assumption 1 (spatial locality) as the weakest point. My stress-test agrees with that identification but sharpens it: the architecture itself may violate the assumption because the multi-scale CNN feature grids provide global receptive fields. This is not merely a theoretical concern; it is likely essential to the method's functioning, since initial poses are sampled uniformly and the score outside the scene cannot be defined. If the score were truly local, a pose in empty space would have zero score and Langevin dynamics would not reliably move samples toward objects. The reported success could therefore be a consequence of global information leakage through low-resolution features, not of the locality prior. The hidden-handle failure mode further supports this: the model sometimes infers handle positions from dataset priors rather than from local geometry around the queried pose. A targeted ablation with a strictly local feature extractor would settle whether the central claim holds. Since the reader's verdict is already CONDITIONAL and this concern is addressable by such an experiment, I recommend keeping the verdict unchanged rather than escalating to rejection.","tokens_in":10839,"tokens_out":4811,"duration_ms":60418,"concrete_test":"Train and evaluate a strictly local variant of SPLIT in which the multi-scale feature grids are computed from the occupancy grid masked to a ball of radius 8 cm around each sample pose before interpolation, or equivalently remove the coarsest global-scale features from Hz. Re-measure packed/pile GSR/DR and mug task success, separating trials where the handle is occluded. If the strictly local variant matches the full model, the locality claim is supported; if performance drops substantially or hidden-handle failures increase, the current results depend on non-local information and Eq. (5) is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central derivation (Eqs. 2-5) requires p(succ|H,X)=p(succ|Hz) and treats p(H|X) as uniform. However, the implemented encoder does not enforce strict locality: Fphi extracts features from multi-scale grids Z1..Z3 produced by five downsampling CNN layers on a 64^3 occupancy grid, so the coarsest grids have receptive fields covering the whole scene. Even a pose query in empty space receives non-zero Hz through low-resolution features, which is likely why Langevin sampling from uniform initial poses can move toward objects. Thus the reported GSR/DR and 70% task success do not cleanly test Assumption 1; they may be explained by global context or dataset priors. The paper's own failure case, where the mug handle is completely invisible in the scene, illustrates a case where success probability depends on a global/shape prior rather than only the local geometry around the sample pose. The claim that score prediction uses 'only local geometric context' is therefore not established by the current experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a new problem formulation, 3D scene-to-pose-set matching, and proposes SPLIT, an SE(3)-diffusion model that generates pose sets from a scene. The method relies on a spatial locality assumption (Assumption 1, Sec. IV-A), under which the success probability of a pose depends only on the local geometric context, leading to the score decomposition in Eqs. (2)-(5). SPLIT encodes local context with a point kernel applied to multi-scale feature grids (Sec. IV-C) and uses sample-frame Langevin dynamics (Sec. IV-D). Experiments evaluate grasp generation on packed/pile scenes (Sec. VI-A) and a multi-purpose mug reorientation/hanging task (Sec. VI-B), reporting GSR/DR and a 70% task success rate versus 77.5% for ground-truth perception.","tokens_in":11035,"tokens_out":5989,"duration_ms":64642,"significance":"The problem formulation is a useful umbrella for grasp detection, object description, and pose-set matching, and the explicit locality-based score decomposition is a clean derivation that could inspire efficient pose-generation models. The paper's strengths are that the core derivation is transparent, the experiments use external task success rather than self-derived metrics, and the mug task demonstrates a single model serving multiple purposes. However, the central claim that score prediction uses only local geometry is not established by the current implementation, because the multi-scale feature encoder has access to global context. The hidden-handle failure in Sec. VI-B shows that the model relies on global priors in exactly the regime where Assumption 1 should be tested. If the locality claim is revised to 'approximately local with learned global priors,' the contribution becomes more modest but still useful.","major_comments":[{"comment":"The implemented geometry encoder is not restricted to local geometry. The 64x64x64 occupancy grid is processed by five downsampling CNN layers, and the feature grids used include coarse scales whose receptive fields span the scene. Since Hz is interpolated from all these scales, a pose query in empty space receives non-zero coarse-scale features, so the network can use global context. This breaks the premise of Eq. (5) in the experiments. Please either enforce locality (e.g., crop a local volume around H, or restrict to fine feature grids) or add an ablation that isolates the contribution of coarse scales and demonstrates that the local part of Hz is sufficient.","section":"Sec. IV-C and Sec. V (Training Details)"},{"comment":"The hidden-handle failure mode is evidence against Assumption 1. When the handle is completely invisible, the local geometry around the successful grasp/hang poses cannot determine the handle position; the model's occasional success comes from dataset priors about mug shape. The paper counts this as a strength, but it directly contradicts the claim that p(succ|H,X)=p(succ|Hz). Please discuss how this case is compatible with Assumption 1, or explicitly position Eq. (5) as an approximate decomposition with global prior terms.","section":"Sec. VI-B, Fig. 5(b)"},{"comment":"Diffusion-EDF [4] is described as the most relevant work but is not included in any comparison. Since the paper claims a general formulation and efficiency advantages over existing SE(3)-diffusion pose generation, a quantitative comparison on the mug task (or at least on a common benchmark) is necessary to support the claim. If a comparison is infeasible, please state the reason and provide a qualitative, but concrete, analysis of expected differences.","section":"Sec. VI-A and Sec. II-C"},{"comment":"The headline result of 70% versus 77.5% task success is based on 40 trials and reported without confidence intervals or statistical tests. At n=40, the standard error is about 7 percentage points, so the observed gap is not significant. Please report confidence intervals (e.g., Wilson intervals) or run more trials, and avoid the phrase '90.3% relative success' without this support.","section":"Sec. VI-B"}],"minor_comments":[{"comment":"The word 'featureugh' appears to be a typo for 'feature through' in the sentence 'based on the concatenated featureugh iterative score-based updates.'","section":"Sec. I (Introduction)"},{"comment":"The notation for gradients is inconsistent: Eqs. (2)-(5) use ∇ without specifying whether it is a global or sample-frame gradient, while the subsequent derivation relies on this distinction. Please define the parametrization once and carry it through.","section":"Sec. IV-B"},{"comment":"The scene capture protocol is described as 'single-view depth camera' in Sec. V, but Sec. VI-B states that 'two depth images were captured from different angles.' Please clarify the protocol for the mug task.","section":"Sec. V and Sec. VI-B"},{"comment":"The point kernel consists of seven points at distance d, but the feature grids have different resolutions; it is unclear whether the kernel displacement d is in physical meters or grid cells, and whether d is tuned. Please specify.","section":"Sec. IV-C"},{"comment":"The SE(3)-Dif baseline omits the signed distance prediction loss, which is a modification of the original method; the effect of this modification on the comparison should be discussed.","section":"Sec. VI-A, Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to be improved by an honest treatment of the locality assumption as approximate. The current framing oversells the theoretical grounding. I would not reject because the experimental approach is externally validated, but the revision must address the architecture's global access and the missing Diffusion-EDF comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper has a genuinely useful idea—predicting SE(3) diffusion scores in the sample's own tangent frame from local point-kernel features—and the experiments show it performing well on a real manipulation task (70% vs 77.5% with ground-truth poses). It deserves a serious referee. But the paper's central theoretical claim, that only local geometry matters, is not actually tested by the current setup, and the evaluation has gaps that need closing.\n\nWhat is new: the sample-frame score parametrization itself, plus the multi-task conditioning that lets one model output grasp, handle, and upright poses from the same scene. That is a worthwhile contribution. The grasp results in Table I beat VGN, GIGA, and SE3-Dif on packed and pile scenes, though the SE3-Dif baseline was modified by omitting its SDF loss. No comparison to Diffusion-EDF, the closest prior work, is a real omission.\n\nWhere the soft spots are, in proportion: the biggest is the mismatch between the locality assumption and the architecture. Equation (5) is valid only if p(succ|H,X)=p(succ|Hz). But the IF-Net-style encoder produces multi-scale grids with global receptive fields after downsampling, so every pose query, even in free space, gets non-zero features from the coarsest grid. That means the model can use global context and scene priors, and the hidden-handle case in Sec. VI-B confirms it does—the paper frames that as a strength, and it is, but it undercuts the claim that local geometry alone drives the score. The 70% task success may come from global priors rather than the proposed local mechanism. This is not fatal to the method, but it means the title and the theory overstate what is demonstrated.\n\nThe multi-task experiment has no error bars, so 70% vs 77.5% is a point estimate from 40 runs. No code is released. The mathematical derivation itself is fine; the score decomposition is a restatement of the assumption, but since the method is validated against external task success, there is no circularity problem.\n\nWho this is for: anyone working on pose generation, grasp detection, or general affordance prediction in manipulation. It is worth a reading-group discussion, especially on the gap between equivariance claims and the translation-equivariant CNN encoder, which the paper honestly flags.\n\nRecommendation: send to peer review. It is a solid paper with a clear idea and plausible results, but the revision should either enforce true locality in the encoder (e.g., restrict the receptive field) or reframe the contribution as localized feature extraction with global context, and add the missing baseline and error bars.","headline":"SPLIT has a genuinely useful architectural idea for SE(3) diffusion, but the locality claim is not cleanly isolated by the experiments and the evaluation needs tightening.","tokens_in":11552,"tokens_out":2604,"would_cite":true,"duration_ms":27041,"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":"SPLIT predicts a pose's success score from only local scene geometry, letting one diffusion model generate grasp, reorient, and hang poses.","keywords":["SE(3)-diffusion","score-based generative model","local geometric context","scene-to-pose-set matching","grasp generation","object reorientation","spatial locality","task conditioning"],"falsifier":"Construct a task where success requires global information — for example, a mug whose handle is entirely outside the 8 cm local kernel, or a placement that must align with a feature on the far side of the scene — and test whether SPLIT's generated poses achieve chance-level success while a model with global context does not; the paper's own reported failure when the handle is completely invisible is a partial instance.","tokens_in":10636,"feed_emoji":"🤖","tokens_out":8878,"duration_ms":83852,"temperature":0.7,"pith_summary":"SPLIT argues that a broad class of robot pose-detection problems — deciding where to grasp, where to place, and which way to orient — can be expressed as one problem: matching a scene to the set of successful poses. The paper's core claim is that the success of a pose depends only on the local geometry around it, so the diffusion score that drives pose generation can be predicted from a local feature vector rather than the whole scene. On this basis, a single model with a task-conditioning vector generates multi-purpose poses for mug reorientation and hanging, reaching a 70% task success rate that is 90.3% of the ground-truth-perception rate (77.5%). The upshot is a perception module that replaces task-specific heuristics with a learned relationship between local geometry and task-relevant poses.","feed_headline":"A single model finds grasp and hang poses from local scene geometry","feed_subtitle":"The same model reorients and hangs mugs at 90 percent of the ground-truth success rate.","key_machinery":"The load-bearing identity is the score decomposition $s(H, X) = \\nabla \\log p(\\mathrm{succ} \\mid H_z)$, which follows from Bayes' rule plus the spatial-locality assumption $p(\\mathrm{succ} \\mid H, X) = p(\\mathrm{succ} \\mid H_z)$ and a uniform prior over poses. The mechanism that realizes it is a local-geometry encoder: a multi-scale convolutional network produces feature grids, a fixed seven-point kernel is transformed by the query pose to sample features near that pose, and the concatenated local features $H_z$ are fed to two MLPs — one predicting the sample-frame SE(3) score, one evaluating sample quality. The sample-frame parameterization, using the local exponential and logarithmic maps on SE(3), makes the score prediction independent of the absolute pose and lets a single encoder serve every pose sample in a scene.","core_discovery":"The paper introduces 3D scene-to-pose-set matching and proves, under the spatial-locality assumption $p(\\mathrm{succ} | H, X) = p(\\mathrm{succ} | H_z)$, that the score function decomposes as $\\nabla \\log p(H \\mid \\mathrm{succ}, X) = \\nabla \\log p(\\mathrm{succ} \\mid H_z)$, so predicting the gradient of log-success from local context is sufficient for SE(3) pose generation. It parameterizes the score in the tangent space of the sample pose, which removes the dependence on the current pose value. The resulting model, SPLIT, encodes local geometry by transforming a small point kernel with the query pose and interpolating a multi-scale convolutional feature grid, then feeds the local vector to a noise-conditioned score network and a sample-evaluation network, both conditionable by a one-hot task vector. Experiments show it matches or exceeds prior grasp-generation baselines in packed and pile scenes, and one model produces the multiple pose types needed for mug reorientation and hanging with task success close to ground-truth perception.","pith_inferences":["If spatial locality holds for other manipulation subgoals, the same architecture could be extended to insertion, pushing, or tool use simply by adding task conditioning and collecting corresponding scene-pose datasets.","The learned score is literally the gradient of log success probability, so SPLIT's output could double as a cost function for motion optimization, connecting pose-sampling with trajectory planning.","A controlled sweep of the point-kernel radius (8 cm in the paper) would quantitatively test how far the locality assumption extends and where global context becomes necessary.","The model's ability to infer a hidden handle from priors suggests that training with more partial views or with a small amount of global context could mitigate its reported failure mode."],"forward_implications":["A single trained model can output different types of poses (grasp, handle direction, upright placement) by switching a conditioning vector, removing the need for separate perception modules per task.","Because the score is predicted from local context and the sample frame, the geometry encoder is shared across all SE(3) samples, giving the architecture a natural translation-equivariance inductive bias.","The model can generate grasps directly from a single-view depth image without object segmentation or problem-specific heuristics.","The mug task result shows that predicted poses reach 90.3% of the ground-truth-perception success rate, leaving a small relative gap from perception alone."],"supporting_citations":[{"why":"Defines SE(3)-diffusion for grasp generation but compresses geometry and pose into one global latent; SPLIT's local score prediction is introduced to overcome this limitation on scene inputs.","marker":"[7]"},{"why":"Supplies the multi-scale convolutional feature grid and point-interpolation approach used to encode local geometric context around a query pose.","marker":"[9]"},{"why":"Provides the packed and pile scene datasets, the declutter task protocol, and the VGN baseline used in the grasp-generation evaluation.","marker":"[14]"},{"why":"Contributes the sample-evaluation network idea, trained with negative and noisy samples, which SPLIT uses to reject low-quality generated poses.","marker":"[18]"},{"why":"Presents the most relevant prior SE(3)-diffusion that consumes full-view scene input with an equivariant encoder, serving as the comparison for SPLIT's lighter local-geometry design.","marker":"[4]"},{"why":"Establishes score-based generative modeling and the score-matching objective that SPLIT's training and sampling are built on.","marker":"[26]"},{"why":"Formalizes denoising score matching, the specific training loss used to fit SPLIT's score network.","marker":"[29]"},{"why":"Defines the SE(3) exponential and logarithmic maps and local parametrization that SPLIT uses for sample-frame score prediction and Langevin dynamics.","marker":"[30]"}],"fun_headline_variants":["One diffusion model finds multiple pose types from local geometry","SE(3) diffusion with local score prediction enables versatile pose matching","Scene to pose-set: one model, local geometry, multiple tasks","Pose prediction from scene geometry, no task-specific hacks","SPLIT diffusion: diverse pose sets from local geometry in one model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that a pose's success probability depends only on the local geometry immediately around that pose — if a successful pose needs information from outside that local neighborhood, such as a hidden handle or a distant constraint, the score decomposition used by SPLIT fails.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion model finds multiple pose types from local geometry","SE(3) diffusion with local score prediction enables versatile pose matching","Scene to pose-set: one model, local geometry, multiple tasks","Pose prediction from scene geometry, no task-specific hacks","SPLIT diffusion: diverse pose sets from local geometry in one model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3066,"prompt_tokens":918,"completion_tokens":2148,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2060}},"tokens_in":534,"tokens_out":2148,"duration_ms":15836,"temperature":1.0,"reasoning_tokens":2060,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:01:29.806115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a task where success requires global information — for example, a mug whose handle is entirely outside the 8 cm local kernel, or a placement that must align with a feature on the far side of the scene — and test whether SPLIT's generated poses achieve chance-level success while a model with global context does not; the paper's own reported failure when the handle is completely invisible is a partial instance.","supporting_citations":[{"cited_title":"Se(3)-diffusionfields: Learning smooth cost functions for joint grasp and motion opti- mization through diffusion,","cited_arxiv_id":null,"evidence_quote":"Defines SE(3)-diffusion for grasp generation but compresses geometry and pose into one global latent; SPLIT's local score prediction is introduced to overcome this limitation on scene inputs."},{"cited_title":"Implicit functions in feature space for 3d shape reconstruction and completion,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-scale convolutional feature grid and point-interpolation approach used to encode local geometric context around a query pose."},{"cited_title":"V olumetric grasping network: Real-time 6 dof grasp detection in clutter,","cited_arxiv_id":null,"evidence_quote":"Provides the packed and pile scene datasets, the declutter task protocol, and the VGN baseline used in the grasp-generation evaluation."},{"cited_title":"6-dof graspnet: Variational grasp generation for object manipulation,","cited_arxiv_id":null,"evidence_quote":"Contributes the sample-evaluation network idea, trained with negative and noisy samples, which SPLIT uses to reject low-quality generated poses."},{"cited_title":"A connection between score matching and denoising autoencoders,","cited_arxiv_id":null,"evidence_quote":"Formalizes denoising score matching, the specific training loss used to fit SPLIT's score network."}],"review_version":1}