{"id":"d3299d03-dce0-4e0e-8d8a-8e62e511ed3d","arxiv_id":"2505.08644","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"DLO-Splatting tracks a deformable rope in 3D by predicting its shape with position-based dynamics and correcting it with a 3D Gaussian Splatting rendering loss.","lead":"A robot rope tracker combines a simple physics model with 3D Gaussian Splatting rendering to estimate a rope's 3D shape from multiple cameras. The approach targets knot tying, where vision alone often cannot tell which strand crosses over which.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 2 in the PBD prediction is dimensionally inconsistent and is not Verlet integration; the paper's central tracking algorithm is not reproducible as written.","rationale":"The reader's weakest assumption targets the predictive adequacy of the position-based dynamics model for self-contact scenarios. My stress-test agrees that prediction is the fragile link, but I identify a more fundamental and concrete problem: Eq. 2 is not a valid integration step as written. The dimensional inconsistency means the algorithm's central component is not reproducible from the text, independent of any empirical evaluation. This is a correctness risk that strengthens the reader's conditional verdict rather than shifting it. I keep the verdict UNCHANGED because the issue might be a correctable typographical omission (e.g., a missing factor of ∆t and mass normalization), and the paper's contribution could survive after a correction and a re-run of the demonstration. The disagreement is partial because the reader identified the prediction as load-bearing but did not flag the equation-level internal inconsistency. I also note the paper's own conclusion admits self-intersections are not modeled, which compounds the problem: even with a corrected integrator, the prediction will not capture topological transitions during knotting, and the single demo shows both methods fail to recover final topology. The concrete test I propose isolates the equation issue and would determine whether the stated algorithm is actually what was run.","tokens_in":8935,"tokens_out":7857,"duration_ms":80795,"concrete_test":"Implement the published Eq. 2 exactly and the corrected version x^{t+1} = 2x^t - x^{t-1} + (F^t/m)∆t^2 (with m set to the node mass and forces from Eq. 3 divided by m), holding all other steps fixed. Run both on the recorded ROS bag used for the Fig. 2 demonstration. Compare the predicted node positions and the final Lobs after the update. If the two trajectories differ by more than a small tolerance (e.g., mean node displacement > 1 rope diameter), the paper's equations do not define the algorithm and the experiment cannot be reproduced from the text. This check would settle whether the inconsistency is a typographical omission or a substantive error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algorithm's prediction step (Section III-A) is the only mechanism that proposes the rope state before the rendering update; it is therefore load-bearing for the central tracking claim. Equation (2) reads X^{t+1} = X^t + (X^t - X^{t-1})/∆t + (1/2) F^t ∆t^2. This is not Verlet integration. The term (X^t - X^{t-1})/∆t has units of velocity and cannot be summed with the position X^t; F^t ∆t^2 carries units of mass-length unless F^t has already been divided by node mass, which the text never states. In fact Eq. (3) defines gravity and friction forces using m_i g, so mass has not been removed. The correct Verlet update for a node of mass m is x^{t+1} = 2x^t - x^{t-1} + (F^t/m)∆t^2, or equivalently x^{t+1} = x^t + v^t ∆t + (F^t/m)∆t^2/2. As written, the integration step mixes dimensions and omits the required factors. Because the PBD output is the initial condition for the Gaussian-splatting update (Eq. 14), any change in the integration formula changes the entire filter. The demonstration's only positive result, tracking the grasped tip, is controlled by the directly-imposed gripper position (Eq. 1), so it does not validate the physics integration. Without a corrected, dimensionally consistent Eq. 2, the claimed prediction capability is not reproducible from the text.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DLO-Splatting, a prediction-update filter for estimating the 3D state of a deformable linear object from multi-view RGB images and gripper pose. The prediction step uses position-based dynamics with gravity, planar contact, friction, and fixed segment-length constraints; the update step renders the predicted node chain with 3D Gaussian splatting and optimizes a rendering loss with SGD. The method is evaluated in one qualitative knot-tying 'cross move' with three cameras and compared against TrackDLO. The authors report that DLO-Splatting tracks the grasped rope tip more accurately than TrackDLO, but that both methods fail to recover the rope's final shape and topology.","tokens_in":9250,"tokens_out":6808,"duration_ms":60304,"significance":"If fully validated, the contribution would be useful: DLO-Splatting combines a training-free physics prediction with a differentiable rendering update, which is a reasonable way to address occlusion and multi-view disambiguation for DLO tracking. The paper also clearly enumerates its limitations, including unmodeled self-intersections, a 1 Hz update rate, and sensitivity to occlusion. However, the central claim that the method tracks visually complex topologies through dense knotting is not supported by the current evidence: Eq. (2) is not reproducible as written, the evaluation is a single qualitative trial without metrics, and the paper itself admits both methods fail on the final topology. The idea merits further work, but the manuscript as submitted is not ready for publication.","major_comments":[{"comment":"Equation (2) is not Verlet integration and is dimensionally inconsistent as written: X^{t+1} = X^t + (X^t - X^{t-1})/Δt + (1/2)F^t Δt^2 adds a velocity-like term to positions, and the force term has units of mass·length unless F^t is already divided by node mass, which Eq. (3) contradicts because the forces include m_i g. The standard Verlet update for a node of mass m_i is x^{t+1} = 2x^t - x^{t-1} + (F^t/m_i)Δt^2, or equivalently x^{t+1} = x^t + v^t Δt + (1/2)(F^t/m_i)Δt^2. This is load-bearing because the PBD output is the initial condition for the rendering update in Eqs. (14)-(15), so any change in the integration formula changes the entire filter. The one qualitative success reported in Section IV (tracking the grasped tip) is controlled by the directly imposed gripper position in Eq. (1), so it does not validate the physics integration. The authors need to provide a corrected, dimensionally consistent integration step and state the mass normalization explicitly.","section":"Section III-A, Eq. (2)"},{"comment":"The evaluation does not support the abstract's and Introduction's claim that DLO-Splatting tracks visually complex topologies that cannot be disambiguated by vision-based tracking alone. The demonstration is a single cross move with three cameras, presented only qualitatively: there are no quantitative metrics such as node-to-ground-truth distance or topology error, no error bars, no ablations, and no sensitivity analysis. The paper's own conclusion states that both DLO-Splatting and TrackDLO fail to recover the rope's final shape and topology, and Section V lists unmodeled self-intersections, a 1 Hz update rate, and occlusion from the gripper as limitations. The only reported advantage, tracking the grasped tip during one move, is too narrow to establish the central contribution, especially because the grasped tip is directly constrained by the gripper pose.","section":"Section IV and Figure 2"},{"comment":"The algorithm is not reproducible from the text because the values of the free parameters used in the demonstration are not reported: friction coefficient μ_f, integration time step Δt, Gaussian diameter σ, node count N and segment length L, Gaussians per segment d, and total rope mass m. The optimizer hyperparameters for the rendering update (learning rate, number of SGD iterations, initialization) are also missing. In addition, the relationship between the constraint projection in Eqs. (4)-(6), which is written as an update to X^t, and the length correction in Eqs. (8)-(9), which is applied to X^{t+1}, is unclear, so a reader cannot reimplement the exact algorithm. Without these details, the single demonstration cannot be checked or extended.","section":"Section III-A and Section IV"}],"minor_comments":[{"comment":"The rendering function h is written as a function of X^t_PBD and the camera projection P_k, but Eqs. (12)-(13) also depend on per-Gaussian colors c_j and opacities o_j; the text should specify how these are initialized and whether they are optimized or fixed during the update.","section":"Section III-B, Eq. (11)"},{"comment":"The velocity v^t_{i,xy} in the friction term is not defined; the text should state whether it is computed from the current and previous node positions and how the planar component is extracted.","section":"Section III-A, Eq. (3)"},{"comment":"Equation (8) is ambiguous without parentheses; it should read Δl^{t+1}_i = (L - l^{t+1}_i) · (x^{t+1}_{i+1} - x^{t+1}_i)/l^{t+1}_i to clarify that the scalar length difference multiplies the unit direction vector.","section":"Section III-A, Eq. (8)"},{"comment":"The gripper action a_t is introduced in the text but does not appear explicitly in Eq. (2); the authors should clarify how the action influences the free nodes as opposed to the grasped node.","section":"Section III-A, Eq. (2)"},{"comment":"The paper should report the image resolution used for the rendering loss, the number of rendering iterations per update, and the actual per-step runtime; Section V mentions a 1 Hz update rate, but the evaluation section does not provide these details.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"This manuscript has the character of an early-stage workshop result: the evaluation is a single qualitative demonstration, the integration step in Eq. (2) is not reproducible, and the conclusion concedes failures on the very topology-tracking task that motivates the work. I believe the correct next step is a major revision with corrected equations, full parameter disclosure, and multi-trial quantitative evaluation; if those cannot be provided, the paper should probably be rejected rather than published in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2505.08644. The combination is genuinely new: position-based dynamics for the prediction step and a 3D Gaussian Splatting rendering loss for the update, applied to ropes. That replaces the learned GNN in Cloth-Splatting, which is a sensible move if you want to avoid training data. The paper is also honest about being early-stage—it lists self-intersection, 1 Hz update, and occlusion as limitations.\n\nThe soft spot is hard to ignore. Equation (2) in Section III-A is not Verlet integration; it's dimensionally inconsistent. It adds a position, a velocity term, and a force term with arbitrary coefficients. The correct Verlet form would be x^{t+1} = 2 x^t - x^{t-1} + (F/m) \\Delta t^2, or velocity Verlet with v \\Delta t. Also the forces are defined with mass, so they need to be divided by mass before use. As written, this prediction step cannot be implemented, and it's the initialization for the Gaussian splatting update. That's load-bearing.\n\nThe demonstration doesn't fill the gap. The one positive result—tracking the grasped tip—comes from Eq. (1), which forces that node to follow the gripper, so it does not validate the dynamics. Both DLO-Splatting and TrackDLO fail on the final shape and topology, and there are no quantitative metrics, ablations, or parameter sensitivity studies. No code or data either. The reader's take of 'conditional' is about right, but I'd go further: the central equation needs correction before the method is reproducible.\n\nNone of this is circular. The rendering loss does provide an independent visual measurement, and the related work is engaged, including self-citations that are appropriate here. The limitations are stated up front.\n\nWho should read this: people working on deformable-object state estimation will find the prediction-update structure and the idea of skipping learned dynamics useful for discussion. As a citation, I wouldn't use it yet. As a paper, I'd send it to peer review—the combination deserves referee time, but the authors need to fix Eq. (2), add real experiments, and release something reproducible.\n\nHope that helps.","headline":"Novel PBD-plus-3DGS combination for DLO tracking, but the central prediction equation is dimensionally off, and the single demo doesn't back the claims.","tokens_in":9841,"tokens_out":3591,"would_cite":false,"duration_ms":34751,"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":"DLO-Splatting tracks a rope's 3D shape by combining physics prediction with Gaussian-splatting image updates.","keywords":["deformable linear objects","3D Gaussian splatting","position-based dynamics","state estimation","knot tying","multi-view RGB","prediction-update filtering"],"falsifier":"Record a full knot-tying sequence in which one segment of the rope is marked with a distinctive color, so the over/under order at each crossing is visible in the images. If, at the moment a crossing is formed, the tracked node chain places the marked segment on the wrong side while the rendered rope still matches all three camera views, then the claim that Gaussian-splatting rendering disambiguates topology is falsified.","tokens_in":8723,"feed_emoji":"🪢","tokens_out":6604,"duration_ms":63129,"temperature":0.7,"pith_summary":"DLO-Splatting estimates the 3D shape of a deformable linear object such as a rope from multi-view RGB images and the gripper pose, using a prediction-update loop that needs no learned dynamics and no fiducial markers. The prediction step evolves a chain of nodes with position-based dynamics under gravity, planar contact, friction, and fixed segment-length constraints. The update step renders the predicted rope as spherical 3D Gaussians, compares the render to each camera image, and adjusts the node positions by stochastic gradient descent on the rendering loss. The demonstration, a cross move used in knot tying, shows the algorithm tracking the grasped rope tip more accurately than a vision-only baseline, though neither method recovers the rope's final shape and topology.","feed_headline":"Physics and Gaussian splatting track rope shape from cameras","feed_subtitle":"A prediction-update filter estimates 3D rope pose during knot tying, no learned dynamics or markers.","key_machinery":"The central machinery is a Bayesian-style prediction-update filter whose state is a node chain $X_t \\in \\mathbb{R}^{N\\times 3}$. The prediction step uses position-based dynamics: Verlet integration with gravity, a planar contact normal, friction, and a length-constraint projection that keeps every segment at length $L$. The update step places $d$ spherical 3D Gaussians per segment along the centerline, renders them into each camera with $\\alpha$-blending, and minimizes the squared image difference $L_{\\text{obs}}$ by stochastic gradient descent on the node positions. The Gaussians' covariance is tied to the rope diameter and their positions to the nodes, so the visual loss directly drives geometric correction of the predicted rope.","core_discovery":"The paper's central claim is that combining a first-principles physics prediction with a differentiable Gaussian-splatting renderer produces a state estimate for a deformable linear object that is robust to visual ambiguities that defeat vision-only trackers, specifically dense crossings and self-occlusion during knot tying. The algorithm represents the rope as $N$ nodes, predicts the next node configuration $\\hat{X}_{t+1}$ with position-based dynamics, then refines it by minimizing the observation loss $L_{\\text{obs}} = \\|\\mathbf{I}^t_k - \\tilde{\\mathbf{I}}^t_k\\|^2$ against all $K$ camera views, where the rendered image $\\tilde{\\mathbf{I}}^t_k$ is computed by $\\alpha$-blending spherical Gaussians placed along the rope centerline. Because the Gaussians are tied to the nodes, the gradient of the rendering loss updates the 3D geometry directly. The authors claim this lets the filter track visually complex topologies that cannot be disambiguated by vision alone, and they demonstrate the advantage on the grasped rope tip during a knot-tying cross move.","pith_inferences":["A direct ablation varying the number of cameras (for example, two vs three) would test whether the reported tip-tracking advantage comes from multi-view disambiguation or from the physics prediction alone; the paper does not report this.","The fixed segment-length position-based dynamics model behaves like a discrete inextensible rod without bending stiffness; comparing against a rod model with bending and twisting energy would isolate whether the final topology collapse is a prediction error or an update error.","The 1 Hz bottleneck is likely shared between the physics time step and the stochastic gradient descent rendering update; a test that runs prediction at 100 Hz while keeping update at 1 Hz would show which side limits the filter.","The claim that vision alone cannot disambiguate these topologies would be sharpened by a controlled experiment with two visually identical crossings, where the only distinguishing information is the camera viewpoint; the paper's single cross-move demonstration is suggestive but not exhaustive."],"forward_implications":["A robot could estimate its rope's 3D state during manipulation using only RGB cameras and its own gripper pose, with no learned dynamics model and no markers.","The same prediction-update loop should transfer to new rope materials or environments without retraining, provided mass, friction, and contact parameters are known.","If the rendering loss truly disambiguates crossing topology, multi-view RGB alone can maintain correct over/under structure where 2D or single-view trackers fail.","The demonstrated 1 Hz update rate and lack of self-intersection modeling currently restrict the method to slow motions and sparse self-contact; faster rendering and explicit self-contact prediction would be needed for full-speed knot tying.","Because the update uses object masking, the method can in principle be extended to simultaneous tracking of several DLOs once multi-instance segmentation is added."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting scene representation and alpha-blending rendering equations that the update step adapts.","marker":"[42]"},{"why":"Provides the vision-based DLO tracking baseline used for initialization and for the qualitative comparison.","marker":"[13]"},{"why":"Supplies the prediction-update cloth state estimation approach with a rendering loss that DLO-Splatting extends to deformable linear objects.","marker":"[12]"},{"why":"Supplies the position-based simulation methods on which the prediction step's constraint projection is based.","marker":"[27]"},{"why":"Supplies the unified particle physics formulation used for the Verlet integration and contact forces.","marker":"[26]"},{"why":"Supplies the fast Gaussian rasterizer used to render the 3D Gaussians into each camera view.","marker":"[47]"}],"fun_headline_variants":["Physics-aware splatting tracks rope knots from multi-view video","Splatting plus physics beats vision-only rope tracking","Gaussian splatting renders rope pose to beat occlusion","First-principles rope tracking with splatting, no markers","Combining physics and splatting for occluded rope tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a simple physics prediction—gravity, table contact, friction, and fixed segment lengths—moves the rope estimate close enough to the true state that the image-based correction can converge, even while the rope crosses over itself during knot tying.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware splatting tracks rope knots from multi-view video","Splatting plus physics beats vision-only rope tracking","Gaussian splatting renders rope pose to beat occlusion","First-principles rope tracking with splatting, no markers","Combining physics and splatting for occluded rope tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1274,"prompt_tokens":871,"completion_tokens":403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":320}},"tokens_in":487,"tokens_out":403,"duration_ms":4355,"temperature":1.0,"reasoning_tokens":320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:49:01.441072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a full knot-tying sequence in which one segment of the rope is marked with a distinctive color, so the over/under order at each crossing is visible in the images. If, at the moment a crossing is formed, the tracked node chain places the marked segment on the wrong side while the rendered rope still matches all three camera views, then the claim that Gaussian-splatting rendering disambiguates topology is falsified.","supporting_citations":[{"cited_title":"3D Gaussian Splatting for Real-Time Radiance Field Rendering,","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting scene representation and alpha-blending rendering equations that the update step adapts."},{"cited_title":"TrackDLO: Tracking Deformable Linear Objects Under Occlusion With Motion Coherence,","cited_arxiv_id":null,"evidence_quote":"Provides the vision-based DLO tracking baseline used for initialization and for the qualitative comparison."},{"cited_title":"Cloth-Splatting: 3D Cloth State Estimation from RGB Supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the prediction-update cloth state estimation approach with a rendering loss that DLO-Splatting extends to deformable linear objects."},{"cited_title":"Position-Based Simu- lation Methods in Computer Graphics","cited_arxiv_id":null,"evidence_quote":"Supplies the position-based simulation methods on which the prediction step's constraint projection is based."},{"cited_title":"Uni- fied Particle Physics for Real-Time Applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the unified particle physics formulation used for the Verlet integration and contact forces."}],"review_version":1}