{"id":"9b69742b-c30a-4b92-9277-d634c15d73bd","arxiv_id":"2502.02663","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A robot estimates an object's center of mass by combining two uncertainty-guided force-torque measurements, showing zero-shot transfer to unseen real-world objects.","lead":"This paper presents U-GRAPH, a system that lets a robot estimate an object's center of mass by taking two force-torque readings while actively choosing a second orientation guided by uncertainty. The method is tested on 12 real-world objects and claims to generalize from training on just two custom objects.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The posterior fusion in Eq. (1) treats two orientation-dependent world-frame CoM estimates as the same quantity; if the outputs are world-frame as stated, no rotation correction is applied, so the second-measurement combination is biased.","rationale":"I read the paper in good faith and find the active perception idea sensible and the experimental evidence promising. The reader's transfer concern is real but partially mitigated by static wrench physics: for a rigid, no-slip grasp, the F/T wrench depends primarily on the object's mass and CoM offset, not directly on its geometry or surface friction, so a network trained on diverse CoM offsets and masses can plausibly transfer. The most load-bearing issue I see is the coordinate-frame inconsistency in the fusion step, because it threatens the internal validity of the central 'second rotation improves accuracy' claim. The paper explicitly says the CoM axes are defined in world coordinates, yet fuses estimates taken at different gripper orientations as if they were the same vector. This is either a missing rotation correction, which would invalidate Eq. (1), or an incorrect definition, which would require rewriting the text and table captions. The current paper does not provide the information needed to decide between these readings. I also note the absence of error bars on the real-world comparison and the mismatch between the abstract's 1.47 cm average and the Table I entries, but those are secondary to the frame issue. My conclusion is that the paper should remain conditional, with the requested clarification and the rotation-consistency check as conditions for acceptance.","tokens_in":9655,"tokens_out":16247,"duration_ms":169049,"concrete_test":"Run a held-out training grasp (or one real object) with known ground-truth CoM offset r. Compute the BNN prediction at orientation [0,0] and at a large tilt theta, e.g., 60 degrees. If mu_2 approximately equals R(theta)*mu_1, the network outputs world-frame vectors and Eq. (1) is inconsistent; re-evaluate after replacing mu_2 with R(theta)^T * mu_2 and check whether errors drop. If instead mu_2 approximately equals mu_1 in the gripper frame, no rotation correction is needed and the 'world coordinates' definition should be corrected. Report the residual under both fusion treatments to settle which interpretation is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A states that dx, dy, dz are defined in world coordinates. Section III-D then describes rotating the gripper to a new orientation, obtaining a second BNN estimate mu_2, and fusing it with the first estimate as mu_final = (mu_1/sigma_1^2 + mu_2/sigma_2^2) / (1/sigma_1^2 + 1/sigma_2^2). If mu is a world-frame CoM displacement relative to the grasp point, then after a rotation R(theta) the same physical CoM offset is R(theta)*r, not r, so mu_1 and mu_2 are estimates of different vectors. Directly averaging them introduces a bias of about (R(theta)-I)*r/2. Since the whole point of the second rotation is to make the z-offset observable, theta is not small, and this bias can be several centimeters. The paper never rotates mu_2 back into the first frame before fusion. If the authors instead intend gripper-frame coordinates, the fusion is valid but then the 'world coordinates' wording and the world-frame axes in Table I are misleading. The reported 1.47 cm headline cannot resolve this ambiguity because the table gives only axis errors without stating the frame used after rotation, and the aggregate number is not reproducible from the table entries.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents U-GRAPH, a two-measurement active perception system for estimating the 3D center of mass of grasped objects. A Bayesian neural network (BNN) maps a 6D force-torque reading and a 2D gripper orientation to a CoM offset and an uncertainty estimate; an ActiveNet, queried by grid search, scores candidate second orientations using the BNN's first-prediction mean and variance, and the system rotates to the best-scoring orientation, takes a second F/T reading, and fuses the two BNN predictions via inverse-variance weighting. The BNN is trained only on two 3D-printed objects with attached laboratory weights, yet the authors report an average error of 1.47 cm (7.6%) on 12 unseen real-world objects and claim zero-shot generalization, with U-GRAPH outperforming one-grasp, analytical, and random-rotation baselines on most objects.","tokens_in":9943,"tokens_out":7663,"duration_ms":76045,"significance":"If validated, this work addresses a genuine and under-explored problem: estimating the 3D CoM of arbitrary rigid objects from haptic interaction, which is useful for stable grasping and manipulation. The use of a Bayesian neural network to produce predictive uncertainty and the formulation of the second rotation as an information-gathering action are reasonable and testable. The evaluation on 12 real objects with independently measured CoM locations, including out-of-distribution weights, is a strength, and the reported OOD degradation in Table II and Sec. V-E is honest. The posterior-combination step is not circular, since the BNN is trained with ground-truth labels and evaluated on unseen objects. However, the frame convention for the fused estimates and the lack of variance reporting prevent me from accepting the quantitative claims as stated.","major_comments":[{"comment":"The fusion in Eq. (1) assumes mu_1 and mu_2 are estimates of the same vector. Section III-A defines dx, dy, dz as world-frame displacements from the grasping point, and the paper does not state that the BNN output for a rotated pose is transformed back to the initial grasp frame before fusion. If the outputs are world-frame, after a gripper rotation R(theta) the physical CoM offset in the world frame is R(theta)*r, not r, and averaging mu_1 and mu_2 without rotating mu_2 back introduces a bias of order (R(theta)-I)*r/2, which can be several centimeters for the large angles used to make the z-offset observable. If the intended convention is that the BNN always outputs the CoM in the initial grasp frame (which the data-collection procedure in Sec. IV-B makes plausible, since labels are computed before rotation), this must be stated explicitly and the training labels must be confirmed to use that convention. As written, the ambiguity is load-bearing for the headline 1.47 cm result and for the comparison with Random Rotate; please state the convention, or apply an explicit rotation correction, and report the results under both conventions.","section":"Sec. III-A, Sec. III-D (Eq. 1)"},{"comment":"Table I reports only per-axis mean errors over five grasps, with no variance or per-grasp breakdown, and the aggregate 1.47 cm / 7.6% accuracy quoted in the abstract is not reproducible from the table. Without error bars or standard errors, the claim that U-GRAPH outperforms Random Rotate cannot be assessed statistically; several entries differ by only a few millimeters. Please report per-grasp 3D errors, standard deviations or confidence intervals, and define the denominator used for the 7.6% figure.","section":"Sec. V-C, Table I"},{"comment":"The analytical baseline formula is written as r = tau × F / |F|^2. The correct inversion of tau = r × F is r = (F × tau) / |F|^2; as written, the formula has the cross-product order reversed. If the implemented baseline used the printed formula, its errors are not meaningful; if it used the correct formula, the text should be corrected. Either way, the baseline comparison needs to be based on the correct formula.","section":"Sec. V-A"}],"minor_comments":[{"comment":"In the Introduction, 'This work provides a framework for perceiving the CoM of an arbitrary object. object.' contains a duplicated word; please fix the typo.","section":"Sec. I"},{"comment":"The citation 'as illustrated in Fig. III-A' should refer to the actual figure number (Fig. 2).","section":"Sec. III-A"},{"comment":"Equation (1) is typeset awkwardly (the fraction is inline and difficult to read); consider displaying it with standard fraction formatting.","section":"Sec. III-D"},{"comment":"The claim of generalization to 'arbitrary rigid object[s]' in the abstract is stronger than the evidence: Table II and the OOD rows of Table I show clear degradation outside the training weight range (43.4 g, 76.8 g, 613.2 g). Please qualify the phrasing, e.g., 'objects within the approximate weight range of the training data'.","section":"Sec. V-E"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the active-perception idea is timely. The main risk is the frame-convention ambiguity in the fusion step; if the authors can resolve it and provide variance estimates, the paper could be acceptable after revision. I would not reject on the basis of the limited training set, as the empirical transfer results are the relevant evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, practically motivated robotics paper with a real experimental contribution, but the fusion step in Sec III-D has a frame-handling problem that needs to be resolved, and the abstract overclaims relative to the paper's own limitations.\n\nWhat's actually new: using BNN uncertainty to actively select a second grasp orientation for 3D CoM estimation, then fusing two F/T measurements. The real-world evaluation on 12 objects is the strongest part — Table I shows U-GRAPH beats the One Grasp, Analytical, and Random Rotate baselines on most axes and objects, and the authors honestly report cases where it loses. The weight-extrapolation study in Sec V-D is good practice.\n\nThe soft spot that matters: Sec III-A defines dx, dy, dz in world coordinates, and Sec III-D fuses mu_1 and mu_2 with inverse-variance weighting. If the outputs are world-frame as stated, then after rotating the gripper the same physical CoM offset becomes R(theta)*r, not r. The paper never rotates mu_2 back before averaging. That's either a mathematical error in the description or a misleading convention statement. The empirical results suggest the implementation might be doing something else, but as written the equations don't close. This needs an explicit fix and clarification, not just a footnote.\n\nSecondary issues: Table I reports only mean errors with no variance across the five grasps, which makes it hard to judge whether the U-GRAPH advantage is reliable. The training set is two custom objects, and the claim of generalization to \"arbitrary\" objects is a stretch given the weight OOD failures the authors themselves report. The abstract says \"remarkable generalizability\" while the limitations section says performance degrades outside the training weight range and with large slips. That's an overclaim.\n\nAlso, the reader's circularity concern doesn't land — the final estimate is supervised prediction on unseen objects, not a re-derivation of training labels. The frame issue is the real technical thing to chase.\n\nWho is this for: roboticists working on physical property estimation and active perception. It deserves a serious referee, not a desk reject. A major revision that fixes the frame consistency, reports variance, and tempers the generalization language would make it acceptable.\n\nMy recommendation: send it to peer review, but flag the fusion-step ambiguity as the first thing the authors must address.","headline":"Useful active-perception CoM paper with real experiments, but the fusion step in Sec III-D has a frame-consistency ambiguity that needs fixing, and the abstract overclaims relative to the limitations.","tokens_in":10434,"tokens_out":3267,"would_cite":false,"duration_ms":32336,"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":"A robot can estimate an unseen object's center of mass by taking two force-torque measurements, with the second orientation picked by a learned uncertainty-guided scorer, reaching 1.47 cm average error after training on just two objects.","keywords":["center of mass estimation","active perception","Bayesian neural network","force-torque sensing","uncertainty quantification","robot manipulation","zero-shot generalization","haptic perception"],"falsifier":"Collect a test set of objects spanning weights from 50 g to 800 g, including low-friction surfaces and shapes unlike the training plate and box, and run both U-GRAPH and the Random Rotate baseline on each. If U-GRAPH does not beat Random Rotate on the z-axis for a majority of objects, or if its average error on objects outside the 127–585 g training range exceeds 3 cm, the zero-shot generalization and uncertainty-guided action-selection claims would be contradicted.","tokens_in":9493,"feed_emoji":"🤖","tokens_out":9578,"duration_ms":80307,"temperature":0.7,"pith_summary":"This paper claims that a robot can estimate the center of mass of an arbitrary rigid object it has never seen before by taking two force-torque measurements at two different grasp orientations. The first measurement feeds a Bayesian neural network that outputs a mean and a standard deviation for the center-of-mass offset; a second network, ActiveNet, scores every candidate rotation angle by how much it would reduce the expected error, and the robot rotates to the best one and measures again. The two estimates are fused as an inverse-variance-weighted average, which the authors show is a simple Gaussian posterior. They report an average error of 1.47 cm (7.6%) on twelve everyday objects, after training on only two custom 3D-printed objects loaded with laboratory weights. If this transfer is reliable, it would give robots a general way to perceive a physical property that is usually assumed known for grasping and manipulation.","feed_headline":"Two guesses beat one: robot finds center of mass of new objects","feed_subtitle":"Uncertainty-guided second rotation reaches 1.47 cm error on unseen real-world objects after training on just two shapes.","key_machinery":"The framework's two load-bearing components are a Bayesian Neural Network (BNN) trained with Markov Chain Monte Carlo using the No-U-Turn Sampler, and ActiveNet, a 5-layer MLP that scores candidate second orientations. The BNN gives each prediction a mean and standard deviation, which is the uncertainty that drives action selection. ActiveNet takes the first estimate's mean and standard deviation plus a candidate second rotation, and predicts the error of the next measurement; a grid search over the continuous 2D orientation space chooses the minimum-score rotation. The final estimate is the inverse-variance-weighted mean of the two Gaussian estimates from the two orientations, treating them as independent measurements.","core_discovery":"The core discovery is that a mapping from 6-axis force-torque readings and grasp orientation to the 3D center-of-mass offset can be learned on a tiny dataset with limited variation and still transfer to arbitrary rigid objects in zero-shot fashion. The model outputs a Gaussian predictive distribution, so the robot knows not only where the center of mass is but how sure it is. That uncertainty is then used to choose a second rotation angle that promises the largest information gain, and the two Gaussian estimates are combined into a posterior. Against one-measurement baselines, the second measurement is what makes z-axis estimation possible, and against a random second rotation, the uncertainty-guided choice consistently yields lower error in the paper's experiments.","pith_inferences":["The success of uncertainty-guided action selection suggests that an ensemble of deterministic networks, which also provides predictive variance, might achieve similar gains at lower training cost; the paper's BNN provides a natural baseline for that comparison.","The inverse-variance fusion treats the two measurements as independent Gaussians, but the second measurement is physically correlated with the first through the shared object; modelling that correlation could improve the posterior and further reduce error.","Because the reported 1.47 cm is an average, its practical value depends on the error distribution; the paper's own data show larger errors on out-of-distribution weights, so deployment should include a weight check or an adapted training set.","The same active-perception loop could be applied to other physical parameters that are observable only through contact, such as friction or compliance, whenever a model can output uncertainty."],"forward_implications":["A single vertical grasp cannot determine the center-of-mass offset along the gravity axis; taking a second, uncertainty-guided rotation always improves the z-axis estimate in the paper's experiments.","Objects whose weights lie inside the training range (127–585 g) are estimated with a few millimeters of error; objects outside that range show noticeably larger errors, so the learned mapping is weight-sensitive.","The trained model transfers to objects with different contact geometry, surface friction, and density, suggesting that the force-torque-to-CoM mapping captures something general about statics rather than memorizing the training shapes.","The two-measurement approach replaces repeated re-grasping with an in-hand rotation, which is faster and safer for delicate objects.","Extending the pipeline to more than two actions, with recursive posterior updates, is the paper's own stated next step and would likely improve accuracy further."],"supporting_citations":[{"why":"supplies the Bayesian framework for neural networks that gives the predictive uncertainty.","marker":"[5]"},{"why":"justifies approximating the infinite ensemble of networks with Monte Carlo sampling.","marker":"[24]"},{"why":"provides the No-U-Turn Sampler used to train the BNN.","marker":"[25]"},{"why":"the Pyro library used to implement the BNN and MCMC.","marker":"[27]"},{"why":"prior work on 2D CoM estimation by iterative pushing, the problem this paper generalizes to 3D arbitrary objects.","marker":"[15]"},{"why":"prior work estimating CoM with torque sensing and reinforcement learning, a baseline context for the new method.","marker":"[17]"}],"fun_headline_variants":["Uncertainty guides robot's second touch to find center of mass","Second guess lowers robot's center-of-mass error","Training on two shapes, robot nails center of mass for new ones","Active perception hones robot's center-of-mass estimate","Uncertainty-guided second rotation beats random choices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mapping between force-torque readings and center-of-mass offsets is learned only from two custom plastic objects with laboratory weights in a narrow range, and the generalization claim assumes that this mapping transfers to arbitrary real-world objects despite differences in geometry, friction, density, and weight.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty guides robot's second touch to find center of mass","Second guess lowers robot's center-of-mass error","Training on two shapes, robot nails center of mass for new ones","Active perception hones robot's center-of-mass estimate","Uncertainty-guided second rotation beats random choices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1838,"prompt_tokens":829,"completion_tokens":1009,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":929}},"tokens_in":445,"tokens_out":1009,"duration_ms":7493,"temperature":1.0,"reasoning_tokens":929,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:31:45.030118+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a test set of objects spanning weights from 50 g to 800 g, including low-friction surfaces and shapes unlike the training plate and box, and run both U-GRAPH and the Random Rotate baseline on each. If U-GRAPH does not beat Random Rotate on the z-axis for a majority of objects, or if its average error on objects outside the 127–585 g training range exceeds 3 cm, the zero-shot generalization and uncertainty-guided action-selection claims would be contradicted.","supporting_citations":[{"cited_title":"A Practical Bayesian Framework for Backpropa- gation Networks,","cited_arxiv_id":null,"evidence_quote":"supplies the Bayesian framework for neural networks that gives the predictive uncertainty."},{"cited_title":"Weight uncertainty in neural networks,","cited_arxiv_id":null,"evidence_quote":"justifies approximating the infinite ensemble of networks with Monte Carlo sampling."},{"cited_title":"Pyro: deep universal probabilistic programming,","cited_arxiv_id":null,"evidence_quote":"the Pyro library used to implement the BNN and MCMC."},{"cited_title":"Predicting center of mass by iterative pushing for object transportation and manipulation,","cited_arxiv_id":null,"evidence_quote":"prior work on 2D CoM estimation by iterative pushing, the problem this paper generalizes to 3D arbitrary objects."},{"cited_title":"Learning and predicting center of mass through manipulation and torque sensing,","cited_arxiv_id":null,"evidence_quote":"prior work estimating CoM with torque sensing and reinforcement learning, a baseline context for the new method."}],"review_version":1}