{"id":"97f7ece4-d961-4052-99b5-bda0a3828b60","arxiv_id":"2502.06109","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Contact Diffusion Model conditioned on force/torque readings and a signed distance field localizes single and dual robot-arm contacts with 0.44 cm and 1.24 cm real-world error.","lead":"The paper introduces a learning-based method that lets a robot arm locate one or two points where it is being touched, using only its own joint and base force sensors. The method is based on a diffusion model, which naturally outputs several possible contact locations instead of one, matching the physical ambiguity of the measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's dual-contact 'error of 1.24 cm' is M-RMSE, a minimum-over-64-samples match; the same table's cluster-center errors are 2.61/2.85 cm, so the headline overstates localization accuracy.","rationale":"I read the paper in good faith and acknowledge its real strengths: a held-out simulation split, an ablation for SDF conditioning, an ablation for historical conditioning, and a real-robot evaluation with labeled markers. The diffusion formulation is a reasonable way to represent the multi-modal posterior induced by sensor singularity, and the paper is honest about many assumptions (point contacts, n_c <= 2, sequential contacts). The reader's stated weakest assumption concerns the point-contact/no-moment model and the frozen first contact. Those are legitimate limitations, but they are explicitly scoped by the paper and do not invalidate the central claim within the stated setting. My stress-test pass identifies a different, more load-bearing issue: the headline accuracy numbers are M-RMSE, a best-of-64 minimum-distance metric, while the same table's cluster-center errors (C-RMSE) are substantially worse. A reader who takes the abstract's '0.44 cm / 1.24 cm' as the error of a localized contact estimate would be misled; the actual mode-center errors are 0.68 cm and about 2.7 cm, respectively. This is not a hidden result - Table III reports both metrics - but the abstract and conclusion highlight the favorable one, and the central claim 'localize contacts with high accuracy' is sensitive to this choice. Because the paper already reports C-RMSE, the issue is primarily one of framing and metric interpretation; it does not require rejecting the method. The reader's CONDITIONAL verdict is still appropriate, so I recommend UNCHANGED. The one concrete check that would settle whether the concern lands is to make C-RMSE (or an equivalent mode-based error) the primary reported metric with per-trial statistics; if those numbers remain as in Table III, the headline claims should be adjusted accordingly.","tokens_in":1236,"tokens_out":944,"duration_ms":100878,"concrete_test":"Re-evaluate the real-world dual-contact data in Table III using the cluster-center error (C-RMSE) or the distance from each k-means mode to the nearest ground-truth contact as the primary metric, reporting per-trial mean and standard deviation. If the mode-based dual error remains approximately 2.6-2.9 cm, as Table III already indicates, revise the abstract and conclusion to state: 'the generated 64-point sample set contains the true contacts within 1.24 cm, while the localized mode centers have error about 2.7 cm,' instead of 'achieves an error of 1.24 cm.' Also report the contact-state classifier's failure rate on the real trials to check if misclassification degrades the historical conditioning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on M-RMSE, defined in Section IV-A as M-RMSE = (1/nc) * sum_j min_{xi in X_t} ||r_t,j - xi||. This is a best-of-64 matching metric: it only asks whether the generated 64-point cloud contains each true contact somewhere, not whether the model's output modes are accurate contact locations. In the real-world results (Table III), the same trials yield C-RMSE-1 = 0.68 cm for single contact and C-RMSE-1/2 = 2.61/2.85 cm for dual contact, i.e., the centers of the k-means clusters are ~2.1-2.3x farther from the true contacts than the abstract's '1.24 cm' figure. Since the stated contribution is to 'localize contacts at arbitrary locations with high accuracy,' the abstract's use of M-RMSE as the headline error materially overstates the method's accuracy for any downstream use that requires a single localized point (which is what 'localization' conventionally means). The method may still be a useful multi-modal sampler, but the quantitative claim as written is not supported by the metric chosen.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDM, a conditional diffusion model for localizing up to two sequential contact points on a robot arm using proprioceptive sensing (joint torque sensors and a base F/T sensor). The model is conditioned on a window of past observations, its own previous single-contact output, and a signed distance field of the robot surface, and it generates a 64-point cloud representing the posterior over contact locations. Training uses simulation data with a 90/10 split; experiments include simulation metrics, ablations on SDF and historical conditioning, and direct sim-to-real transfer on a Kinova Gen3 arm with 100 single-contact and 100 dual-contact trials. The abstract claims real-world localization errors of 0.44 cm (single) and 1.24 cm (dual) at 15.97 ms inference time.","tokens_in":1769,"tokens_out":1810,"duration_ms":60343,"significance":"If the quantitative claims are properly interpreted, CDM is a novel and promising application of generative modeling to contact localization: it is, to my knowledge, the first diffusion-based contact localizer, it handles multi-modality explicitly, and the real-robot dual-contact evaluation is a useful addition to a literature that has mostly evaluated in simulation. The 90/10 held-out simulation split, the direct real-robot transfer test, and the ablation studies are genuine strengths. The main weakness is that the headline accuracy metric is not the error of a localized point, so the claimed 'high accuracy' overstates what the system delivers for downstream point-estimate tasks.","major_comments":[{"comment":"The headline real-world errors (0.44 cm single, 1.24 cm dual) are M-RMSE values, defined as (1/nc) * sum_j min_{xi in X̂_t} ||r_t,j - xi||. This metric only asks whether each true contact has at least one nearby sample among the 64 generated points; it does not report the error of any localized point estimate. The same trials in Table III give C-RMSE-1 = 0.68 cm for single contact and C-RMSE-1/2 = 2.61/2.85 cm for dual contact, i.e., the cluster-center errors are roughly 2–2.3 times larger than the abstract's figures. Since 'localization' conventionally means producing a specific point, and downstream controllers need a point estimate, the abstract's use of M-RMSE as 'the error' materially overstates the method's accuracy. Please report C-RMSE (or another pointwise error) as the headline accuracy, and state explicitly what M-RMSE does and does not certify.","section":"Section IV-A, Table III, Abstract"},{"comment":"The real-world results in Table III report only mean values over 100 trials, with no standard deviation, confidence interval, or distribution information, despite the robot configuration being randomized every 20 experiments. Table II likewise gives no variance for the 614,400 simulation runs. Without dispersion measures, the reader cannot assess the reliability of the claimed 'high accuracy' or the significance of the 40% improvement over the PF baseline [12]. Please include variance statistics or, at minimum, report per-trial error distributions (e.g., median and interquartile range) for the real-world experiments.","section":"Table III and Table II"},{"comment":"Algorithm 2 freezes X̂_Ts once the contact state classifier labels the current observation as single-contact, and it never re-estimates X̂_Ts during subsequent dual-contact phases. This design assumes the first contact remains stationary after it appears. If the first contact slides or rolls during the dual-contact phase, or if the classifier mislabels a steady dual-contact as single, the historical conditioning is based on a stale first-contact location. The assumptions in Section III.A (point contact without moment, up to two sequential contacts) are stated, but the consequence of first-contact motion is not discussed or evaluated. Please add an explicit limitation statement and, ideally, an experiment with a moving first contact to delimit the range of validity of the central claim.","section":"Algorithm 2 and Section III.A"},{"comment":"The paper's central contribution is the prediction of a multi-modal posterior, yet no quantitative metric evaluates whether the generated distribution is calibrated or whether the number and location of modes are correct. M-RMSE is a coverage metric that ignores spurious modes, and C-RMSE forces k-means clustering with the number of clusters equal to the true contact count, which cannot detect extra or missing modes. Since the singularity-handling claim rests on the model producing the correct multi-modal structure, please add distribution-level evaluation, such as mode count accuracy, coverage of high-probability regions, or a proper scoring rule for the empirical posterior.","section":"Section IV-A, Fig. 4, Table III"}],"minor_comments":[{"comment":"The section title contains a typo: 'Denosing diffusion probabilistic models' should read 'Denoising diffusion probabilistic models.'","section":"Section II.C"},{"comment":"The loss in Algorithm 1 is written as Lθ = ||ϵ - ϵθ(...)||, omitting the square that appears in Eq. (6); please make the two consistent.","section":"Algorithm 1"},{"comment":"The dataset notation D = {ξ_i}_{N_d}^{i=1} is unconventional; it should be written as D = {ξ_i}_{i=1}^{N_d}.","section":"Section III.B"},{"comment":"Footnote 1 correctly notes that direct comparison with [20] and [21] is complicated by the additional base F/T sensor, but the main text still states 'the proposed method achieves higher accuracy than existing learning-based methods.' Please qualify this comparison in the main text as well, not only in the footnote.","section":"Section IV.D, Table III"},{"comment":"The SDF network fφ is described as pre-trained and taken from [30], but no details are given about its training data, resolution, or whether it is fine-tuned during CDM training. Please clarify, since the SDF conditioning is a central architectural component.","section":"Section III.D / Fig. 3"},{"comment":"The contact-state classifier failure rates are reported per state, but the label definition for 'trans-dual' (nc = 1 → 2) versus 'steady dual' (nc = 2) is not explicit in Section III.B. Please define precisely what labels gψ is trained to predict for each of the three states.","section":"Table I"},{"comment":"The equation for M-RMSE is given inline without a number, which makes it awkward to reference in the text and in tables; please number it.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a robotics conference/journal and the core idea is sound. The main issue is that the abstract's headline accuracy numbers use M-RMSE, a minimum-over-64-samples coverage metric, which substantially overstates the point-localization accuracy; this needs to be fixed by changing the reported headline metric or by clearly relabeling it. There is no sign of circularity: the evaluation uses held-out simulation data and a real-robot transfer test. The self-comparison to [12] is a legitimate baseline, though the metric difference should be handled more carefully. I recommend major revision rather than rejection because the underlying method and experiments are credible, and the concerns are addressable by re-analysis and clarification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the important thing to know: this is a legitimately new application of diffusion models to contact localization, and the real-robot dual-contact results are unusual. The core trick—conditioning the denoiser on past model outputs and an SDF—works well in their experiments. The paper is well structured, with a proper held-out simulation split and a sim-to-real transfer test. I don't see circularity or fitting to the test set. The self-citation to their prior PF work is a fair baseline.\n\nThe main thing to flag is the metric. The abstract's '1.24 cm' dual-contact error is M-RMSE, which measures whether any of the 64 generated points lands close to each true contact. That's a coverage or containment metric, not a single-point localization error. The same real-world trials give cluster-center errors of 2.61 and 2.85 cm. For anyone using this as a 'localization' module to get a point estimate, the abstract is misleading. The paper does report C-RMSE in the tables, so it's not hidden, but the headline overstates accuracy for single-point use.\n\nOther soft spots: no variance measures for the 100 real-world trials, no code or data released, and the method assumes pure point contacts (no moment), a fixed first contact once classified as single, and at most two sequential contacts. These are stated, so they're limitations rather than hidden flaws. The simulation only uses synthetic contacts with friction-cone forces; real-world contacts might be less clean.\n\nNone of this kills the contribution. As a multi-modal sampler for contact distribution, CDM is a step forward. It handles the singularity problem that sinks particle filters, runs at 16 ms, and the ablations show both the SDF and the historical conditioning matter. The paper deserves a serious referee, and I'd engage with it. For a review, I'd ask for cluster-center metrics in the abstract or at least an honest footnote, variance across runs, and ideally code/data release to make the sim-to-real reproducibility checkable.","headline":"A genuinely new diffusion-based approach to multi-contact localization with real-robot evidence; the headline error metric flatters the method, but the core contribution is solid.","tokens_in":12293,"tokens_out":1733,"would_cite":true,"duration_ms":15544,"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":"The paper's central claim is that a diffusion model conditioned on joint torques, base wrench, past outputs, and a signed distance field can localize one or two contact points, including cases where several contact/force pairs produce…","keywords":["contact localization","diffusion model","multi-contact","proprioceptive sensing","signed distance field","robot manipulation","multi-modal posterior","denoising diffusion probabilistic model"],"falsifier":"Run the trained model on a trajectory where a single contact slides along the arm after the classifier has frozen the first-contact estimate; if the frozen $\\hat{X}_{T_s}$ does not track the sliding point, the fixed-first-contact assumption fails and the real-world M-RMSE should exceed the reported 1.24 cm dual-contact error.","tokens_in":11326,"feed_emoji":"🤖","tokens_out":10536,"duration_ms":80551,"temperature":0.7,"pith_summary":"CDM treats contact localization as posterior sampling rather than point estimation: given a window of joint-torque and base-wrench observations, the model draws 64 candidate contact points from the distribution conditioned on those observations and on the signed distance field of the robot's surfaces. The paper's central claim is that this generative formulation resolves the singularity in which multiple pairs of contact points and forces produce exactly the same sensor readings, because the samples can cover all modes instead of collapsing to one. A second claim is that conditioning on the model's own earlier output—freezing the first contact once a classifier says only one contact is present—reduces multi-modality when a second contact appears sequentially. If these claims hold, multi-contact localization no longer needs the threshold tuning that particle-filter methods require, and one model handles single-, transition-, and steady-dual-contact cases. The paper reports real-world matched RMSE of 0.44 cm for single contact and 1.24 cm for dual contact, at 15.97 ms per inference.","feed_headline":"Diffusion model pinpoints robot contacts to 0.44 cm","feed_subtitle":"When two contacts produce identical torque readings, it outputs all plausible contact points instead of one guess.","key_machinery":"The load-bearing object is the conditional denoising diffusion model with a FiLM (feature-wise linear modulation) conditioned MLP denoiser. At each reverse step $k$, the denoiser predicts the noise given a window of proprioceptive observations, the saved first-contact estimate $\\hat{X}_{T_s}$ (or null when no first contact is known), the diffusion step $k$, and, for each candidate point, the signed distance and gradient to every link surface from a pretrained SDF network. These conditions enter through FiLM layers: one per-point affine transform from the SDF feature and one shared affine transform from a global observation feature, so the surface geometry directly reshapes the denoising trajectory. A separate contact-state classifier, trained with cross-entropy loss, decides whether the current window is single-contact; if it is, the historical first contact is frozen to the current output. The training loss is the standard DDPM noise-prediction loss, with an extra random diffusion step applied to the first contact point so the network learns with and without historical information.","core_discovery":"The discovery is that the ill-posed mapping from contact points and forces to sensor measurements, which leaves many equally consistent explanations, is better handled by sampling the posterior than by optimizing a single estimate. CDM uses a denoising diffusion network whose reverse process is conditioned on a window of proprioceptive observations, the current joint configuration, and the signed distance field (SDF) of every link, and it outputs a set of points that jointly cover the true contact locations. The contact-state classifier identifies whether the current window is single-contact; if so, the first-contact estimate is saved and used as an additional condition for subsequent windows, so a sequentially appearing second contact is localized relative to an already-fixed first contact. In simulation the method reaches M-RMSE (distance from each true contact to the nearest generated sample) of 0.32 cm for single contact and 1.14 cm for steady dual contact; in real-world tests it reaches 0.44 cm and 1.24 cm, respectively, using ten denoising steps at 15.97 ms.","pith_inferences":["An implication the authors leave implicit: the same posterior-sampling formulation could be extended to contact-force identification, because the measurement model is linear in the forces once the contact geometry is sampled; the paper lists this as future work.","A testable extension beyond the paper: because the model transfers from simulation to hardware without fine-tuning, its checkpoints could be evaluated on a different arm of similar kinematics after recomputing the SDF, which would separate the method's generality from its training distribution.","An editorial observation: the reported single-contact accuracy relies on an extra base F/T sensor that earlier learning-based comparisons may lack, so the 0.44 cm figure likely mixes the value of the generative model with the value of that additional sensing modality.","A possible consequence for control: the 64-sample posterior could be summarized by its modes and fed to contact-aware model-predictive control, turning the multi-modal distribution into robust constraint sets rather than a single assumed contact."],"forward_implications":["One model covers single-contact, transition dual-contact, and steady dual-contact states, eliminating the user-tuned threshold used by particle-filter methods to decide when a new contact appears.","Because the output is a set of samples, singular cases with several valid contact positions remain represented instead of collapsing to one point estimate, so downstream planning can act on the full uncertainty.","The 15.97 ms inference time for ten denoising steps is constant regardless of contact count and is fast enough for closed-loop contact-aware control.","SDF conditioning keeps generated points on the robot surface: the average distance between generated samples and the surface is 0.29 cm with SDF versus 0.89 cm without.","Historical conditioning more than halves M-RMSE and cluster-center error in sequential dual-contact localization compared with null conditioning (M-RMSE 0.62 cm versus 1.33 cm in simulation)."],"supporting_citations":[{"why":"provides the recursive particle-filter baseline that CDM compares against and whose threshold-based new-contact initialization CDM aims to remove.","marker":"[12]"},{"why":"introduces the contact particle filter that keeps particles on the robot surface and motivates using past observations for sequential contacts.","marker":"[16]"},{"why":"supplies the denoising diffusion probabilistic model formalism and the simplified noise-prediction training loss.","marker":"[24]"},{"why":"gives the momentum-based disturbance observer used to estimate external joint torques for the observation window.","marker":"[27]"},{"why":"extends the disturbance observer to estimate the base wrench used in the conditioning window.","marker":"[28]"},{"why":"provides the signed-distance-field network used to condition denoising on robot surface geometry.","marker":"[30]"},{"why":"supplies the deterministic denoising schedule used for ten-step inference.","marker":"[37]"},{"why":"is one of the learning-based single-contact baselines whose reported error the real-world result is compared against.","marker":"[20]"}],"fun_headline_variants":["Diffusion model pinpoints robot contacts to 0.44 cm","Ambiguous touches resolved by diffusion sampling: 0.44 cm","Sampling beats guessing: robot contacts to 0.44 cm","Contact diffusion nails multi-touch to 0.44 cm real world","Diffusion handles torque ambiguity, localizes contacts to 0.44 cm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that there are at most two contacts, each is a pure point contact transmitting no moment, the first contact stays fixed once detected, and the momentum-observer wrench estimate is unbiased; if any of these fails, the conditioned posterior and the frozen historical contact are wrong.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model pinpoints robot contacts to 0.44 cm","Ambiguous touches resolved by diffusion sampling: 0.44 cm","Sampling beats guessing: robot contacts to 0.44 cm","Contact diffusion nails multi-touch to 0.44 cm real world","Diffusion handles torque ambiguity, localizes contacts to 0.44 cm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1371,"prompt_tokens":916,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":361}},"tokens_in":532,"tokens_out":455,"duration_ms":4889,"temperature":1.0,"reasoning_tokens":361,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:43:10.734803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on a trajectory where a single contact slides along the arm after the classifier has frozen the first-contact estimate; if the frozen $\\hat{X}_{T_s}$ does not track the sliding point, the fixed-first-contact assumption fails and the real-world M-RMSE should exceed the reported 1.24 cm dual-contact error.","supporting_citations":[{"cited_title":"Proprioceptive sensor-based simultaneous multi-contact point localization and force identification for robotic arms,","cited_arxiv_id":null,"evidence_quote":"provides the recursive particle-filter baseline that CDM compares against and whose threshold-based new-contact initialization CDM aims to remove."},{"cited_title":"Localizing external contact using pro- prioceptive sensors: The contact particle filter,","cited_arxiv_id":null,"evidence_quote":"introduces the contact particle filter that keeps particles on the robot surface and motivates using past observations for sequential contacts."},{"cited_title":"Collision detection and safe reaction with the dlr-iii lightweight manipulator arm,","cited_arxiv_id":null,"evidence_quote":"gives the momentum-based disturbance observer used to estimate external joint torques for the observation window."},{"cited_title":"Combining real and virtual sensors for measuring interaction forces and moments acting on a robot,","cited_arxiv_id":null,"evidence_quote":"extends the disturbance observer to estimate the base wrench used in the conditioning window."},{"cited_title":"Contact point localization for articulated manipulators with proprioceptive sensors and machine learning,","cited_arxiv_id":null,"evidence_quote":"is one of the learning-based single-contact baselines whose reported error the real-world result is compared against."}],"review_version":1}