{"id":"3c2fb96d-5646-4e52-9598-138d1948b9fa","arxiv_id":"2505.00500","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"INR-DOM reconstructs a complete implicit surface of elastic bands from partial point clouds and fine-tunes the representation with contrastive RL, achieving up to 89 percent simulated and 90 percent real-world manipulation success.","lead":"This paper trains a robot to handle stretchy rubber bands by first learning to imagine the band's full 3D shape from a partial camera view, then using that internal picture to guide reinforcement learning. The approach reports markedly higher success than existing methods in three simulated tasks and around 80-90 percent success on a real robot arm, which matters for robotic tasks with soft, easily deformed objects.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world claim rests on unvalidated SAM2 segmentation: Section VII admits the dependence, but no robustness experiments quantify how segmentation errors affect the reported success rates.","rationale":"The reader's weakest_assumption identifies the segmentation dependency, and this is also the most load-bearing concern. The real-world half of the central claim depends on a segmentation pipeline that the paper itself acknowledges as critical, yet the paper offers no robustness study. The simulation results are not affected by this concern, so the overall verdict remains CONDITIONAL rather than ACCEPT or REJECT: the method may perform well when segmentation is accurate, but the real-world claim requires an explicitly validated perception component. I considered other objections—missing error bars and significance tests, unreleased code, and minor inconsistencies in reported percentages—but those weaken the evidence without identifying a specific mechanism by which the claim would fail. The segmentation dependency is concrete, acknowledged, and untested, so it is the appropriate focus. No change to the reader's CONDITIONAL verdict is needed.","tokens_in":14650,"tokens_out":4883,"duration_ms":58092,"concrete_test":"Run the existing Franka setup for sealing, installation, and disentanglement under at least three conditions: standard lighting, strong directional lighting, and a partial occlusion of the band. Manually label each SAM2 mask as correct or incorrect, compute mask IoU against a hand-labeled mask, and record task success. If success drops below 70% when mask IoU remains above 0.8, or if any occlusion condition produces frequent mask failures, the real-world claim is conditional on segmentation and must be reported with that quantified caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that INR-DOM significantly improves success rate for DOM tasks—is supported in simulation by clean, simulator-generated partial point clouds, but the real-world experiments in Section V.C feed the same PointNet encoder with point clouds produced by a SAM2-based segmentation and outlier-removal pipeline running at 10 Hz. Section VII explicitly states: 'The performance of INR-DOM depends on the accuracy of the point-cloud segmentation model.' Yet the paper provides no experiment that varies segmentation quality, occlusion, lighting, or twist complexity, and no analysis of how segmentation errors propagate to latent embeddings and actions. Because the encoder is pre-trained on partial/complete pairs with clean segmentation, and because no mechanism in the policy can detect or recover from a corrupted mask, the reported real-world success rates (90% over 10 trials; 80% in the harder comparison) are conditional on a component that is never validated. If SAM2 fails on heavily twisted or occluded bands, the downstream representation and policy have no way to recover, so the real-world transfer claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes INR-DOM, a two-stage representation-learning and control method for manipulating elastic deformable objects (rubber bands and O-rings) from partial point-cloud observations. Stage I pre-trains a PointNet-based encoder as a partial-to-complete variational autoencoder whose hypernetwork decoder predicts the weights of a SIREN signed-distance-field network, trained with SDF, skeleton, KL, weight, and consistency losses. Stage II fine-tunes the encoder jointly with soft actor-critic and an InfoNCE contrastive loss that uses temporal and instance-wise key selection via start-goal embedding distance and dynamic time warping. The paper reports reconstruction accuracy on nine simulated band types, success rates on sealing, installation, and disentanglement tasks in IsaacSim, learning curves, t-SNE latent visualizations, and real-robot trials on a Franka Emika Panda arm with SAM2-based segmentation.","tokens_in":14899,"tokens_out":4729,"duration_ms":49172,"significance":"The paper addresses a relevant and timely problem in deformable-object manipulation. If the results hold, the contribution is useful: it demonstrates that reconstruction-based pre-training with implicit neural SDFs can be combined with contrastive RL fine-tuning to yield manipulation policies from partial observations. The leave-one-out reconstruction evaluation, the ablations of pre-training and contrastive components, and the real-robot comparison against SERL are commendable and give direct evidence for the main message. However, the statistical support for the word 'significantly' is missing, and the real-world transfer claim rests on an unvalidated segmentation component and an unvalidated image-based reward classifier. With additional experiments or carefully scoped claims, the work could be a solid contribution; in its current form, the central claims are stronger than the evidence presented.","major_comments":[{"comment":"The paper repeatedly states that INR-DOM 'significantly improves convergence stability in policy learning and success rate', but Table I reports single success rates per method with no confidence intervals, error bars, or number of seeds, and Fig. 8 shows one learning curve per method. For 100 Bernoulli trials, the 85% sealing result has a 95% confidence interval of roughly 76-91%, so some pairwise differences with the strongest baselines are clear, but the central claims about convergence stability and statistical significance are not supported by the reported evidence. Please report means and standard deviations over at least five random seeds and, ideally, a paired test across environments, or soften the wording to 'improves' without the statistical claim.","section":"Section V.A, Table I and Fig. 8"},{"comment":"The paper acknowledges in Section VII that 'The performance of INR-DOM depends on the accuracy of the point-cloud segmentation model' and that the real system uses SAM2 segmentation plus outlier removal at 10 Hz, yet no experiment quantifies robustness to segmentation errors, occlusion, lighting changes, or severe twisting. Because the encoder is pre-trained on clean simulator partial clouds and no mechanism is described for detecting or recovering from a corrupted mask, the reported real-world success rates (90% over 10 trials per task, and 80% in Table II) are conditional on a component that is never validated. Please add a segmentation-perturbation study in simulation (e.g., random mask erosion or dilation, missing points, added outliers) that measures the effect on end-to-end task success, or explicitly re-scope the real-world claim to 'given accurate segmentation'.","section":"Section V.C and Section VII"},{"comment":"The real-world experiments use a learned image-based reward classifier that assigns positive reward when the current RGB image 'closely matches' the final scene from expert demonstrations, but the paper reports no accuracy or failure analysis for this classifier. Since the policy is optimized against this learned reward, the reported success rates may partly reflect the classifier's ability to recognize goal images rather than genuine task completion, and the absence of any validation of the reward classifier is a load-bearing gap for the real-world claim. Please report classifier accuracy, a confusion matrix, or a task-completion check independent of the learned reward.","section":"Section IV.B and V.C"}],"minor_comments":[{"comment":"The InfoNCE loss is written as the positive log-ratio rather than its negative; as printed, minimizing the expression would maximize agreement between query and positives, contradicting the stated goal. Please correct to L = -log(...) for reproducibility.","section":"Equation (6)"},{"comment":"There is a typo: 'porint cloud' should be 'point cloud'.","section":"Section III.A"},{"comment":"The phrase 'pre-trained segmentation anything model 2 (SAM2)' should be 'pre-trained Segment Anything Model 2 (SAM2)'.","section":"Section IV.B"},{"comment":"The statement 'an average rate 40.3% higher than ACID' should specify that this is in percentage points, not a relative percentage improvement, to avoid ambiguity.","section":"Section V.A"},{"comment":"The repeated column headers 'Model Rate [%]' are confusing; the table would be clearer with a single 'Method' column and a single 'Success rate [%]' column.","section":"Table II"},{"comment":"The label '2·10^4 embeddings' would be easier to read as '20,000 embeddings'.","section":"Fig. 7 (b)"}],"recommendation":"major_revision","confidential_remarks":"I would be willing to review a revised version. The reconstruction-oriented pre-training is well supported, but the strength of the RL and real-world claims is not yet matched by the evidence. Adding multi-seed statistics and a segmentation-robustness study would materially change my assessment. The paper's scope fits the journal well, and I do not see a novelty disclosure problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on 2505.00500. This is a useful empirical paper, not a breakthrough, but it earns its central claim directionally, and the ablations are good enough that both stages of the pipeline look load-bearing.\n\nWhat's actually new: the system assembles known pieces (PointNet-style encoder, hypernetwork-decoded SIREN SDF, SAC, CURL-style InfoNCE) but the assembly is new, and two components stand out — the skeleton loss for intertwined bands and the DTW-based temporal/instance key assignment for the contrastive loss. The ablation table is the strongest part of the paper: dropping pre-training collapses success to 20/29/16, and dropping contrastive fine-tuning gives 58/61/54 versus 85/89/75, so the design choices matter. Reconstruction beats PCN, PoinTr, and Point2Vec on both CD and EMD, including on unseen band sizes. The simulation margins over ACID are large and plausible.\n\nSoft spots, in rough order. (1) No error bars, confidence intervals, or significance tests on the headline numbers. The word \"significant\" is doing work it hasn't earned. The margins are wide enough that this is fixable, but the report is incomplete as written. (2) The real-world evaluation is thin. Ten trials per task is a small foundation for a 90% claim, and the system's dependence on SAM2 segmentation — acknowledged in Section VII but never stress-tested — means the real-world numbers are conditional on a component the paper doesn't validate. The stress-test note is right that heavy occlusion or twist could break segmentation with no recovery mechanism in the policy. The SERL comparison is also confounded by camera modality (RGB versus depth), so it doesn't cleanly isolate the representation. (3) Reproduction: several hyperparameters (loss weights, EMA momentum, number of negatives, M) are unreported and no code is released.\n\nNone of this sinks the core claim. The simulation results are coherent, the ablations are honest, and the limitations section reads like the authors know where the weak points are. Whoever reads this will get a clear picture of what is contribution and what is contingency.\n\nWho it's for: people working on deformable-object manipulation, representation learning for RL, or sim-to-real transfer for point-cloud policies.\n\nRecommendation: send it to peer review. A referee should push for error bars, a segmentation-robustness experiment (adding synthetic noise to the partial point cloud would be cheap), and a larger real-world trial. The paper deserves the referees' time.","headline":"Solid, honest empirical paper on implicit representations for elastic deformable-object manipulation; large simulation margins and convincing ablations, but real-world claims rest on 10 trials and an unvalidated SAM2 segmentation dependency.","tokens_in":15441,"tokens_out":4530,"would_cite":true,"duration_ms":43115,"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":"INR-DOM shows that pretraining a PointNet encoder to reconstruct full signed-distance shapes from partial point clouds, then fine-tuning it with contrastive reinforcement learning, yields state representations that push elastic-band…","keywords":["deformable object manipulation","implicit neural representation","signed distance function","point cloud completion","contrastive learning","reinforcement learning","elastic bands","occlusion-robust representation"],"falsifier":"A single decisive experiment would run the real disentanglement task with the crossing of the +180/-180 degree twist deliberately occluded in the segmented point cloud, for example by a pole blocking the camera's view; if task success stays high, the representation is genuinely occlusion-robust, and if success collapses to baseline levels, the reported rates depend on segmentation quality rather than on the SDF representation.","tokens_in":14461,"feed_emoji":"🤖","tokens_out":8543,"duration_ms":84465,"temperature":0.7,"pith_summary":"This paper argues that a robot can manipulate elastic bands it cannot fully see if the state representation is built from a complete implicit shape model rather than from raw or discretely completed point clouds. INR-DOM first pre-trains a PointNet encoder to turn a partial point cloud into a latent code that reconstructs the band's full signed distance field, then fine-tunes that encoder with reinforcement learning and a contrastive loss that pulls temporally matched configurations together. The paper reports that this two-stage recipe raises task success to 85% (sealing), 89% (installation), and 75% (disentanglement) in simulation, an average 40.3 percentage points above the next-best baseline, and around 90% per task on a real robot arm. A sympathetic reader would care because deformable-object manipulation has resisted compact state representations: infinite degrees of freedom and self-occlusion make both reconstruction and policy learning hard, and this is a concrete proposal for making the representation do both jobs.","feed_headline":"One neural SDF representation lifts elastic-band robot success by 40%","feed_subtitle":"Reconstructing the band's full shape from partial point clouds gives reinforcement learning a stable state space that transfers to a real…","key_machinery":"The central object is the implicit signed distance function (SDF): a neural network that takes a query point and a latent code and returns the signed distance to the band's surface, so the complete geometry is encoded continuously even when the input point cloud is partial. The mechanism that carries the argument is a two-stage training loop. Stage one is a partial-to-complete variational autoencoder: a PointNet encoder produces a 64-dimensional latent code, a hypernetwork decodes that code into the weights of a SIREN SDF network, and reconstruction, skeleton, and consistency losses force the code to capture the full shape. Stage two fine-tunes the encoder with soft actor-critic and an InfoNCE contrastive loss whose positive keys come from a temporal- and instance-wise selection scheme based on episode similarity and dynamic time warping, so the latent space separates entangled configurations that look similar in raw observations. The SDF is doing the key work: it supplies a continuous, complete geometric scaffold that discrete point-cloud completion methods do not provide.","core_discovery":"On its own terms, the paper's discovery is that representing a partially observed elastic band by its complete signed distance function, decoded by a SIREN hypernetwork from a 64-dimensional latent code, gives reinforcement learning a state space that is both dense and occlusion-tolerant. The pre-training loss combines surface reconstruction, medial-axis (skeleton) preservation, a consistency loss that forces partial and complete embeddings to match, and standard variational and weight regularizers. In the fine-tuning stage the same encoder is updated jointly with a soft actor-critic policy and an InfoNCE contrastive loss, where positive keys are chosen by dynamic time warping across similar episodes rather than by naive anchor-augmentation pairs. The result, as reported, is a latent space in which twist direction is encoded consistently, with opposite twist angles landing on opposite sides while completed trajectories converge to the goal region, and the learned policy transfers to a real robot with 20 demonstrations, reaching roughly 90% success per task.","pith_inferences":["Beyond the paper's claims, the same partial-to-complete SDF pretraining should extend to other deformable linear objects such as cables and ropes whenever a medial-axis ground truth can be computed, since the skeleton loss is the component most tied to band-like topology.","The paper does not test segmentation robustness, so a fair extension would be to feed full depth images and let an attention mechanism focus on the band; the stated limitation implies that doing so would remove the main reported failure mode.","The temporal key-assignment recipe, using episode similarity plus dynamic time warping to pick positives, is a general technique for any off-policy reinforcement learning agent with sequential observations, not just elastic bands, and could be evaluated on cloth smoothing or rigid-object pushing.","A testable consequence the authors leave implicit is that INR-DOM should also predict the medial axis of unseen band types well, because Lskel is a training signal; measuring medial-axis error directly would isolate whether the skeleton loss or the SDF loss drives generalization."],"forward_implications":["If the reported results are correct, a single pretrained shape-completion encoder can replace hand-designed state features for a family of elastic-band tasks, because the same latent codes support sealing, installation, and disentanglement.","The contrastive fine-tuning is load-bearing for the hardest task: removing it drops disentanglement success from 75% to 54%, so the temporal key assignment is what lets the policy distinguish +180 degrees from -180 degrees twists.","The reconstruction pretraining is indispensable: without it, success rates fall to 20-29%, indicating that random initialization of the encoder cannot be recovered by reinforcement learning alone.","The same representation is expected to generalize to unseen band sizes and types, since reconstruction error on unseen bands is lower than the baselines' error on seen bands."],"supporting_citations":[{"why":"Supplies the sinusoidal-activation SIREN network whose weights the hypernetwork decoder predicts for the implicit SDF.","marker":"[37]"},{"why":"Supplies the latent-conditioning scheme for predicting signed distance fields from an embedding.","marker":"[28]"},{"why":"Provides the soft actor-critic algorithm whose actor, critic, and entropy losses define the RL fine-tuning objective.","marker":"[10]"},{"why":"Provides the template for joint contrastive representation learning and reinforcement learning fine-tuning.","marker":"[14]"},{"why":"Defines the InfoNCE contrastive objective used to pull temporally matched states together and push unrelated states apart.","marker":"[27]"},{"why":"Supplies the momentum-updated key encoder used to generate positive and negative keys for contrastive learning.","marker":"[12]"},{"why":"Provides the base PointNet encoder architecture that INR-DOM modifies.","marker":"[30]"},{"why":"Defines the ACID baseline, the next-best implicit-representation DOM method that INR-DOM claims to beat by 40.3% on average.","marker":"[35]"},{"why":"Defines the Point2Vec point-cloud completion baseline that INR-DOM outperforms on reconstruction error.","marker":"[1]"}],"fun_headline_variants":["Implicit neural fields give robots a handle on elastic bands","Neural SDFs make deformable-object RL more robust","Reconstructing the full shape of an elastic band improves robot RL","Full shape from partial views boosts elastic-band RL","Implicit surfaces give RL a stable state for elastic bands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the point cloud entering the encoder is a clean, correctly segmented view of the target band; the real-world system relies on SAM2-based segmentation plus outlier removal, and the paper does not validate what happens when occlusion, lighting, or severe twisting breaks that segmentation.","fun_headline_variants_meta":{"raw":{"variants":["Implicit neural fields give robots a handle on elastic bands","Neural SDFs make deformable-object RL more robust","Reconstructing the full shape of an elastic band improves robot RL","Full shape from partial views boosts elastic-band RL","Implicit surfaces give RL a stable state for elastic bands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000936,"raw_usage":{"total_tokens":3992,"prompt_tokens":920,"completion_tokens":3072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2990}},"tokens_in":536,"tokens_out":3072,"duration_ms":22582,"temperature":1.0,"reasoning_tokens":2990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:40:31.333155+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A single decisive experiment would run the real disentanglement task with the crossing of the +180/-180 degree twist deliberately occluded in the segmented point cloud, for example by a pole blocking the camera's view; if task success stays high, the representation is genuinely occlusion-robust, and if success collapses to baseline levels, the reported rates depend on segmentation quality rather than on the SDF representation.","supporting_citations":[{"cited_title":"Implicit neural representations with periodic activation functions","cited_arxiv_id":null,"evidence_quote":"Supplies the sinusoidal-activation SIREN network whose weights the hypernetwork decoder predicts for the implicit SDF."},{"cited_title":"Deepsdf: Learning continuous signed distance functions for shape representa- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the latent-conditioning scheme for predicting signed distance fields from an embedding."},{"cited_title":"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor","cited_arxiv_id":null,"evidence_quote":"Provides the soft actor-critic algorithm whose actor, critic, and entropy losses define the RL fine-tuning objective."},{"cited_title":"Curl: Contrastive unsupervised representations for rein- forcement learning","cited_arxiv_id":null,"evidence_quote":"Provides the template for joint contrastive representation learning and reinforcement learning fine-tuning."},{"cited_title":"Momentum contrast for unsupervised visual representation learning","cited_arxiv_id":null,"evidence_quote":"Supplies the momentum-updated key encoder used to generate positive and negative keys for contrastive learning."},{"cited_title":"Guibas, Silvio Savarese, Anima Anandkumar, and Yuke Zhu","cited_arxiv_id":null,"evidence_quote":"Defines the ACID baseline, the next-best implicit-representation DOM method that INR-DOM claims to beat by 40.3% on average."},{"cited_title":"Point2vec for self-supervised repre- sentation learning on point clouds","cited_arxiv_id":null,"evidence_quote":"Defines the Point2Vec point-cloud completion baseline that INR-DOM outperforms on reconstruction error."}],"review_version":1}