{"id":"6d5f5a2b-bcf0-412a-8f29-d4b87fabc1e4","arxiv_id":"2506.12184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SPLATART builds Gaussian splat renderers that can redraw articulated objects in new configurations, using sparse part segmentations and estimating joint structure for deep kinematic chains.","lead":"SPLATART learns articulated object models from posed photos and a few labeled part images by separating part discovery from joint estimation. The pipeline is demonstrated on the Paris dataset, a real trigger clamp, and simulated robot arms with up to seven joints.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Two-scene identifiability gap: any joint with zero or near-zero motion between the only two observed configurations cannot be uniquely estimated, and the paper provides no analysis or test of this necessary condition for the central claim.","rationale":"The reader's weakest assumption correctly identifies the two-scene sufficiency as the most fragile premise. My stress-test sharpens this into a concrete identifiability condition: every joint must exhibit non-zero, well-conditioned relative motion between the two observed configurations. The paper gives no analysis of this condition and no evidence that the robot-arm experiments were chosen to satisfy it. This is not a minor implementation detail; it directly affects whether the estimated joint parameters and the resulting configuration-conditioned renderer are correct for all joints. The paper's positive results on the Paris dataset and the qualitative/limited quantitative robot-arm results are promising, but they do not resolve this gap. The appropriate verdict remains CONDITIONAL, pending an identifiability analysis or an experiment with a static joint. Since my review does not change the reader's conditional verdict, I recommend UNCHANGED.","tokens_in":9433,"tokens_out":7265,"duration_ms":95895,"concrete_test":"Render two scenes of the Panda URDF (or any 7-DOF arm) where one joint (e.g., joint 4) has the same angle in both scenes while all other joints move. Run SPLATART with the same settings as Section IV-C. If the estimated axis position and angle for the static joint deviate significantly from ground truth (e.g., rotation error > 1° or axis position error > 1 cm), or if varying that joint in the learned model produces no correct motion, then the two-scene assumption is insufficient. Repeat with a small angle change (e.g., 0.5°) to test conditioning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SPLATART estimates per-part poses and joint parameters for tree-structured articulated objects (including 7-DOF robot arms) presupposes that every joint's motion is observable in the two scenes used. Section III-A states 'this work assumes only two scenes are present'. For a revolute joint, the axis is determined by the relative rotation between the two observed part poses; if the joint angle is identical in both scenes (or differs by a very small amount), the relative rotation is near-identity and the axis position is ill-conditioned or completely unidentifiable. The same holds for a prismatic joint with zero displacement. The paper neither states this identifiability condition nor reports whether the Panda/Fanuc experiments used configurations where all joints moved sufficiently. Consequently, the claimed ability to render arbitrary user-specified configurations is not supported for static or near-static joints. The joint estimation procedure in Section III-D fits revolute/prismatic candidates from pairwise part poses and accepts them via a threshold; with a zero-motion joint, multiple candidates can pass or the axis is degenerate, yielding an incorrect representation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SPLATART is a pipeline that builds articulated 3D Gaussian splat renderers from posed RGB images with part segmentation masks. It first trains per-scene semantic Gaussian splats, then jointly optimizes per-part rigid poses across two observed configurations using a differentiable rendering loss, estimates pairwise revolute and prismatic joint candidates by minimizing an ADD-style distance on part means, assembles a kinematic tree by depth-first search, and renders novel configurations by transforming the canonical part splats. The paper reports quantitative comparisons on the Paris dataset against PARIS and DTA, a qualitative real-world trigger-clamp experiment with sparse segmentations, and experiments on simulated Franka Panda and Fanuc serial-chain manipulators, with joint-state errors for the Panda shown in Table III.","tokens_in":9667,"tokens_out":8695,"duration_ms":96634,"significance":"The proposed decoupling of geometry/pose estimation from articulation estimation is a sensible design and, if properly supported, would be a useful contribution to articulated-object representation for robotics. The paper also targets deeper kinematic chains than most prior work on articulated NeRFs/Gaussian splats, and it evaluates joint predictions against external ground truth (Paris metrics, URDF joint errors), so the pipeline is not circular. The real-world sparse-segmentation demonstration is a valuable qualitative step. However, the current manuscript is not yet publishable: an entire table is empty, the robot-arm evaluation has no comparison baseline and no axis-parameter metrics, the two-scene formulation has an unaddressed identifiability problem, and the reported quantitative results lack error bars. These gaps directly affect the central claims, but they appear fixable within the scope of the paper.","major_comments":[{"comment":"The two-scene formulation makes joints with zero or near-zero relative motion between t=0 and t=1 unidentifiable or ill-conditioned, because the relative part pose is the only cue used to fit the revolute and prismatic candidates. The paper neither states this necessary identifiability condition nor reports the joint displacements used in the Panda and Fanuc experiments. Since the central claim includes rendering arbitrary user-specified configurations for a 7-DOF arm, the authors should add an identifiability analysis and report per-joint motion magnitudes (or add a third observation) to demonstrate that every estimated joint was actually exercised.","section":"Section III-A / III-D"},{"comment":"Table II, captioned \"Quantitative results for the relative joint states,\" is empty and is never referenced in the text, so the promised quantitative evaluation of joint-state estimation is missing. This is not a presentation detail: the abstract's claim that \"we present data on the SPLATART pipeline as applied to the Paris dataset objects\" depends on this table. The authors must either fill the table with the corresponding metrics or remove the claim and revise the text accordingly.","section":"Table II (Section IV-A)"},{"comment":"The robot-arm evaluation does not support the strength of the central claim. No comparison to any prior articulated-object method is provided; Table III reports only joint-state errors, not errors in the estimated joint axes or axis positions, which are the quantities needed to support the claim that SPLATART recovers the articulation model. Figure 5 only interpolates the estimated t=0 and t=1 configurations along a one-parameter path, so it does not demonstrate rendering at arbitrary user-specified 7D configurations. Table III also lists 9 joints while the text says \"up to seven degrees of freedom\"; the authors should clarify the URDF variant and whether the gripper joints are included.","section":"Section IV-C / Table III"},{"comment":"All quantitative results in Table I are single-run values with no error bars, significance tests, or seed information. Several entries are close to or worse than the DTA baseline (e.g., USB axis angle 4.61 degrees vs 0.11 degrees, fridge CD-m 20.35 mm vs 0.27 mm), so the claim of \"results comparable to previous radiance field based efforts\" in the conclusion is not statistically supported. The authors should report mean and standard deviation over multiple runs or otherwise justify the stability of the pipeline.","section":"Table I (Section IV-A)"},{"comment":"The conclusion's claim that SPLATART works \"without a prior knowledge of the object structure\" is overstated: the method requires part segmentation masks, which encode the number of parts and part membership, and the problem formulation takes N_p as given. Additionally, no experiment with a branching kinematic tree is reported; all Paris objects are single-DOF two-part objects and the robot arms are serial chains, so the tree-structure generation step (Section III-E) is not validated for branching topologies. The claims should be qualified accordingly or evaluated on a branching object.","section":"Section III-B / V"}],"minor_comments":[{"comment":"There are many typos, including \"syntheic,\" \"spare\" (for \"sparse\"), \"seperation,\" and \"exmaples\"; a careful proofread is needed before resubmission.","section":"Abstract and throughout"},{"comment":"The loss vector uses \"lossrgb\" and \"lossssim\" while Eqs. (1)-(4) define \"lossl1\" and \"lossss\"; the notation should be aligned.","section":"Eq. (5)"},{"comment":"The acceptance threshold epsilon and the loss weights lambda_* are described as user-specified, but their values are not reported anywhere in the manuscript or supplement; these values are needed for reproducibility.","section":"Section III-D"},{"comment":"The axis-position rows for \"Blade\" and \"Storage\" show dashes for PARIS; the authors should explain whether these metrics are undefined or omitted.","section":"Table I"},{"comment":"The caption is confusing (\"The top set of three rows are the poses for time t=0\" in a four-quadrant layout); the panels should be relabeled and explained more clearly.","section":"Figure 4"},{"comment":"Reference [19] has no venue information; complete the citation.","section":"References"},{"comment":"The real-world clamp experiment is qualitative only; a quantitative metric (e.g., reprojection error or joint-angle error) would strengthen the claim of good real-world geometric and kinematic results.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early draft: an empty table and uncorrected typos in the abstract are serious editorial issues. The main technical concerns are the two-scene identifiability gap, the missing robot-arm baseline and axis-parameter metrics, and the lack of error bars in Table I. These are fixable within the scope of the paper, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SPLATART is a pipeline integration rather than a new principle, but it does something genuinely useful: it decouples part segmentation from joint estimation and shows the approach on seven-joint robot arms, which prior articulated-object work rarely touches. If the results hold up, it lowers the cost of building articulated models for manipulation and simulation.\n\nWhat is new and what works: The explicit two-stage design - learn semantic Gaussian splats, get per-part poses across two scenes, then fit joints post hoc - is a sensible separation that makes the articulation model modular. The demonstrated extension to deep kinematic trees (Panda and Fanuc arms) is real and, as far as I know, not present in the cited baselines. The sparse-segmentation real-world clamp demo is a good proof of concept. Table III provides quantitative joint errors for the Panda, and they look reasonable.\n\nThe soft spots are mostly about evidence, not architecture. Table II is referenced but empty; that's a clear production error. Table I reports single-run values with no error bars, so we can't judge stability. The robot-arm experiments have no comparison to prior articulated-object methods like GART or ArtGS, which weakens the claim of advantage. The real-world experiment is qualitative only. No code is released.\n\nThe more substantive concern, flagged by our stress-test, is the two-scene assumption. Section III-A states only two scenes are used. For any joint that does not move between those two scenes, the relative pose is near-identity and the axis position is ill-conditioned or unidentifiable. The paper does not discuss this condition or report whether the arm experiments used configurations where all joints moved enough. This is a load-bearing gap for the claim of arbitrary-configuration rendering. It is fixable - run an experiment with a static joint, or report joint motion ranges and an identifiability analysis - but it should be addressed.\n\nOverall, the paper is worth engaging with. The deep-tree demonstration is a contribution that a serious referee can help shape. I recommend accepting for peer review, with the expectation of revisions: fill Table II, add error bars and baselines for the arm results, and either justify or qualify the two-scene identifiability issue.","headline":"Useful pipeline integration with a real deep-tree demonstration, but the evidence is thinner than the claims and the two-scene identifiability issue needs addressing.","tokens_in":10164,"tokens_out":2284,"would_cite":true,"duration_ms":26682,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SPLATART estimates part poses, joint parameters, and connectivity for tree-structured articulated objects, including seven-degree-of-freedom robot arms, from posed RGB images with sparse part segmentations, and renders configurations…","keywords":["articulated objects","Gaussian splatting","kinematic tree estimation","part pose estimation","joint parameter estimation","robot manipulators","sparse segmentation","novel configuration rendering"],"falsifier":"Capture a seven-joint arm in two scenes with one joint intentionally held nearly still (less than about one degree of motion), run the pipeline, and compare the estimated axis for that joint against the other joints; if the still joint's axis is not recovered accurately, the two-scene assumption is falsified.","tokens_in":9248,"feed_emoji":"🦾","tokens_out":7425,"duration_ms":85735,"temperature":0.7,"pith_summary":"This paper argues that the articulation structure of a tree-structured object—which parts there are, how they are posed in each observed scene, and which joints connect them—can be recovered from posed RGB images where only a subset of frames carry part segmentations, without any prior kinematic model. The proposed pipeline, SPLATART, builds a Gaussian splat for each scene, attaches semantic part scores to the splats, then jointly optimizes per-part poses and splat parameters across the two scenes so that each part renders correctly in the other scene's camera views. Once part poses are known, joint candidates are fit and a kinematic tree is assembled from pose differences. The paper demonstrates this on a synthetic two-part benchmark and on simulated robot arms with up to seven revolute joints, reporting part-motion errors at or below a strong baseline and rendering intermediate configurations that were never observed during training. A sympathetic reader would take the central bet to be that two observed configurations plus sparse labels are enough to synthesize an articulated object's full configuration space.","feed_headline":"Two posed scenes build a reconfigurable 3D robot arm","feed_subtitle":"SPLATART recovers joint axes and part poses from sparse labels, then renders arm poses never seen in training.","key_machinery":"The load-bearing object is the part-conditioned Gaussian splat: each rigid part is a set of 3D Gaussians with means, covariances, opacities, and spherical harmonics, plus a learned vector that scores how strongly each Gaussian belongs to each part. The mechanism that carries the argument is cross-scene pose optimization: for every part and every scene pair, the estimated transform between the two scenes is used to move that part's Gaussians into the other scene's coordinate frame, where they are rasterized and compared against the other scene's RGB, accumulation, and segmentation images. Because the splat parameters and the poses are updated together during this rendering loss, the geometry of each part is forced to explain both observed configurations; this is what produces completed geometry and accurate part motion. The joint estimation that follows is deliberately decoupled from geometry: revolute joints are parameterized by a rotation center, axis, and a fixed transform, prismatic joints by a fixed transform and translation axis, and a candidate joint is accepted only if transforming the source part's splat means yields an average distance to the target part's splat means below a threshold.","core_discovery":"The central claim is that a configuration-conditioned renderer for an articulated object can be learned without knowing the object's structure in advance, using only two posed scenes and partial part labels. SPLATART treats each rigid part as its own canonical Gaussian splat with a semantic score per splat, and represents each scene by a pose per part. The part poses and the splat geometry are optimized together by rendering every part into the other scene's frames and minimizing a combined RGB, accumulation, segmentation, and SSIM loss, which both aligns the motions and fills in geometry that was occluded in one scene. After the poses are fixed, a separate step fits revolute and prismatic joints between every pair of parts using an ADD distance threshold and builds the final kinematic tree by depth-first search from a chosen root. This separation of geometry from articulation is what the paper credits for scaling to deeper kinematic trees, and it reports that on a seven-joint robot arm the per-joint state errors stay mostly under one degree.","pith_inferences":["Editorial inference: the two-scene assumption means the method is only as good as the motion captured between the scenes; an object with a joint that barely moves between t=0 and t=1 will have that joint's axis underdetermined, so a natural extension is to add more scenes until every joint has moved enough.","Editorial inference: because tree generation is a depth-first pass over accepted pairwise joints, the same pipeline could be extended to closed kinematic loops by replacing the tree builder with a graph solver, which the paper explicitly leaves to future work.","Editorial inference: the separation of articulation from geometry suggests a testable upgrade path where part segmentations are produced automatically rather than manually, which would make the whole pipeline nearly label-free while keeping the same cross-scene pose machinery."],"forward_implications":["From two posed scenes with sparse part labels, the pipeline produces a renderer that can synthesize configurations never seen in training, including intermediate joint states.","Articulation structure—joint type, axis, and connectivity—is recovered without a prior model, so the same machinery applies to tools, cabinets, and seven-joint robot arms.","Because joint estimation is a separate post-processing step, changing the joint model (for example to screw or spherical joints) does not require retraining the splat.","Jointly optimizing poses and splat geometry lets each part borrow visual evidence across scenes, so occluded regions of the object get completed automatically.","Real-world data collected with sparse manual labels (on the order of tens of images per scene) is enough for qualitative novel-configuration rendering."],"supporting_citations":[{"why":"Provides the two-part articulated object benchmark and the comparison metrics used in the main experiments.","marker":"[1]"},{"why":"Defines the 3D Gaussian splatting representation that the pipeline extends with semantic part scores.","marker":"[5]"},{"why":"The simulation renderer used to generate synthetic RGB images and part masks for the benchmark and robot arm experiments.","marker":"[8]"},{"why":"The main baseline whose axis angle, axis position, part motion, and Chamfer distance numbers SPLATART is compared against.","marker":"[9]"},{"why":"Source of the serial-chain manipulator models used to demonstrate deep kinematic tree estimation.","marker":"[10]"},{"why":"The pose-based kinematic model estimation work that motivates making joint estimation a separate post-processing step.","marker":"[11]"},{"why":"Supplies the per-scene Gaussian splat optimization routine used at the start of the pipeline.","marker":"[20]"},{"why":"Provides the observation that coupling pose and scene-parameter optimization avoids local minima during pose refinement.","marker":"[24]"},{"why":"Defines the ADD matching score used as the loss for fitting and accepting joint candidates.","marker":"[25]"}],"fun_headline_variants":["Two scenes, sparse labels: SPLATART infers robot joints","Learn articulated structure from only two posed views","Gaussian splats + two scenes = estimated robot kinematics","Infer joint axes and part poses from two scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on two observed configurations of the object being enough to determine every joint, and on the supplied part segmentations covering the object's full geometry.","fun_headline_variants_meta":{"raw":{"variants":["Two scenes, sparse labels: SPLATART infers robot joints","Learn articulated structure from only two posed views","Gaussian splats + two scenes = estimated robot kinematics","Infer joint axes and part poses from two scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000861,"raw_usage":{"total_tokens":3747,"prompt_tokens":966,"completion_tokens":2781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":2715}},"tokens_in":582,"tokens_out":2781,"duration_ms":24111,"temperature":1.0,"reasoning_tokens":2715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:56:36.857813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Capture a seven-joint arm in two scenes with one joint intentionally held nearly still (less than about one degree of motion), run the pipeline, and compare the estimated axis for that joint against the other joints; if the still joint's axis is not recovered accurately, the two-scene assumption is falsified.","supporting_citations":[{"cited_title":"Paris: Part-level reconstruc- tion and motion analysis for articulated objects,","cited_arxiv_id":null,"evidence_quote":"Provides the two-part articulated object benchmark and the comparison metrics used in the main experiments."},{"cited_title":"Sapien: A simulated part-based interactive environment,","cited_arxiv_id":null,"evidence_quote":"The simulation renderer used to generate synthetic RGB images and part masks for the benchmark and robot arm experiments."},{"cited_title":"Neural implicit representation for building digital twins of unknown articulated objects,","cited_arxiv_id":null,"evidence_quote":"The main baseline whose axis angle, axis position, part motion, and Chamfer distance numbers SPLATART is compared against."},{"cited_title":"Understanding urdf: A dataset and analysis,","cited_arxiv_id":null,"evidence_quote":"Source of the serial-chain manipulator models used to demonstrate deep kinematic tree estimation."},{"cited_title":"A probabilistic framework for learning kinematic models of articulated objects,","cited_arxiv_id":null,"evidence_quote":"The pose-based kinematic model estimation work that motivates making joint estimation a separate post-processing step."},{"cited_title":"Parallel inversion of neural radiance fields for robust pose estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the observation that coupling pose and scene-parameter optimization avoids local minima during pose refinement."},{"cited_title":"Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes,","cited_arxiv_id":null,"evidence_quote":"Defines the ADD matching score used as the loss for fitting and accepting joint candidates."}],"review_version":1}