{"id":"2104267b-13e5-4016-b13b-cd048b3f1c48","arxiv_id":"2505.00647","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A projection-based force estimator and ellipsoidal-uncertainty planner enable dexterous grasping and extrinsic manipulation with inaccurate tactile force readings.","lead":"GeoDEx is a framework that lets robots plan and control grasps and tool pivoting using noisy force readings from tactile fingertips. It improved grasp success in hardware trials and ran roughly 14 times faster than a standard optimization baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The robust-planning contribution rests on an unvalidated Gaussian noise model and is never ablated, so its role in the reported success is untested.","rationale":"The reader's conditional verdict is appropriate. My concern aligns with the reader's weakest assumption but sharpens it: even accepting the empirical success rates, the paper does not show that the uncertainty-aware planning component is what enables the improvement. The Gaussian noise model in Definition 5 is internally contradicted by the sensor characterization in Section IV-A and by the Discussion's own admission. Without reported σ_i values and without an ablation isolating the planner, the advertised robustness under force uncertainty is not established. This does not invalidate the hardware results, but it does mean the framework's central novelty is currently untested. A conditional acceptance that requires an ablation (or, if the ablation fails, an empirical validation of the ellipsoid) would settle the matter, so I keep the reader's verdict unchanged.","tokens_in":14069,"tokens_out":7149,"duration_ms":73557,"concrete_test":"In MuJoCo, replicate the wrench and cylinder grasping experiments using the identical estimator and admittance controller but replace the robust planner (Problem 3) with a nominal planner that sets σ_i = 0 in the same linear constraints. Run 20 trials per condition per the paper's protocol and compare success rates and force errors. If nominal planning matches the robust planner's roughly 80% and 60% success rates, the ellipsoidal uncertainty model is not the source of robustness. If it degrades significantly, then report the σ_i values actually used and verify against the ATI F/T sensor that the empirical error distribution is contained within the 1σ ellipsoid at the claimed confidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"GeoDEx's central claim is that its geometric planner and estimator are robust to inaccurate tactile readings. That robustness guarantee rests on Definition 5, which assumes per-contact normal-force errors are independent and Gaussian with known variances σ_i, so the trusted region is an ellipsoid. Section IV-A, however, characterizes the Touchlab sensors as having activation-threshold jumps (0.1–0.5 N), hysteresis offsets, partial-contact errors, and taxel-dependent noise, none of which are Gaussian or independent. The paper never reports the σ_i values used in the hardware experiments, and Section V admits that modeling the error \"is extremely difficult, if not impossible.\" Thus the ellipsoid's size and shape are unspecified, and the feasibility constraints in Problems 3 and 4 do not provide a tested guarantee against the actual error distribution. Moreover, the hardware comparison always uses the full system (estimator plus robust planner plus admittance controller) against raw measurements; there is no ablation that replaces the robust planner with a nominal planner while keeping the estimator. The observed success-rate improvements may therefore be due entirely to the FE-plane projection in the estimator, not to the uncertainty-aware planning that is the paper's headline contribution. If that is the case, the theoretical framework is not load-bearing, and the claim that GeoDEx handles force uncertainty is overstated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeoDEx, a geometric framework for force estimation, robust force planning, and admittance control for tactile-based dexterous and extrinsic manipulation. The core idea is to represent force-equilibrium contact forces as a hyperplane (FE-plane), represent noisy normal-force measurements as a cone (M-Cone), and model measurement uncertainty as an ellipsoid on the FE-plane; planning then becomes a linear feasibility problem (Problems 3 and 4) that is reported to be 14x faster than SOCP. Hardware experiments on an Allegro hand with Touchlab tactile fingertips show that using the estimator improves grasp success from 20% to 80% for a wrench and from 0% to 60% for a cylinder, and keeps cube-pivot angular error under 5 degrees, compared with raw tactile readings. Simulation results show convergence of estimated forces and successful cube-turning.","tokens_in":14307,"tokens_out":4818,"duration_ms":43936,"significance":"If the robustness guarantee held as stated, the paper would make a useful contribution by providing a computationally cheap geometric alternative to SOCP for force planning under tactile uncertainty and by demonstrating that noisy tactile readings can be used for closed-loop force control in dexterous and extrinsic manipulation. The hardware comparisons are a genuine strength: they involve repeated trials, report success rates and force errors, and directly compare estimator-based feedback with raw measurements. The framework is not fitted to reproduce the experimental outcomes; the geometric projections and constraints are derived from stated assumptions, with only noise variances and object parameters as inputs. However, the paper does not establish that the robust-planning component is load-bearing, because no ablation isolates it, and the Gaussian ellipsoid model is not validated against the sensor characterization presented in Section IV-A. As a result, the theoretical guarantee in Problems 3 and 4 is currently a stated assumption rather than a demonstrated result.","major_comments":[{"comment":"The trusted-measurement ellipsoid assumes per-contact normal-force errors are independent and Gaussian with known variances, but this contradicts the paper's own sensor characterization (activation-threshold jumps of 0.1-0.5 N, hysteresis offsets of 0.0-0.2 N, partial-contact and angular errors of 0.1-0.5 N, and taxel-dependent noise), and Section V states that modeling the error 'is extremely difficult, if not impossible.' The sigma_i values used in the hardware experiments are never reported, so the ellipsoid's size and shape are unspecified and the feasibility guarantee of Problems 3 and 4 is not tested against the actual error distribution. The authors should report sigma_i, justify the Gaussian assumption, or validate robustness under distribution mismatch.","section":"Sec. III-C, Definition 5 and Problems 3-4 vs. Sec. IV-A"},{"comment":"All experimental comparisons are between the full system (estimator plus planner plus admittance controller) and raw measurements; there is no ablation that keeps the estimator but replaces the robust planner with a nominal planner. Since the estimator is the only component that is claimed to be evaluated by the 'switching feedback' experiments, the reported success-rate improvements could be entirely due to the FE-plane projection in Sec. III-B rather than the uncertainty-aware planning that is the paper's headline contribution. An ablation with a nominal planner is needed to establish that the robust-planning formulation is load-bearing.","section":"Sec. IV-B and IV-C"},{"comment":"The objective for extrinsic force estimation is not clearly derived. The term ||B_sub f_est|| is presented as the distance to the measurement sub-space cone, but B_sub is defined as a basis of that cone, so its geometric meaning is unclear; the summation indices i and ne in the constraint f_est = P_fe * sum_i w_i f^i_sub are inconsistent; and the relationship between this QP and the 'minimum magnitude' rationale in the text is not formalized. Please provide a complete derivation and correct the notation.","section":"Sec. III-B, Eq. (7)"},{"comment":"The simulation uses a tactile sensing noise variance of 0.5 N, but the hardware characterization reports per-taxel steady-state noise of only 0.01-0.03 N and per-taxel errors of 0.1-1 N. The paper does not state how the 0.5 N value is derived from the taxel-level characterization or what sigma_i is used for the hardware experiments, so the connection between the simulated and hardware uncertainty models is missing. Please specify the mapping.","section":"Sec. IV-B and IV-A"}],"minor_comments":[{"comment":"The heading contains a typo: 'Perservation of force equilibrium' should be 'Preservation of force equilibrium.'","section":"Sec. III-A, Proposition 2 heading"},{"comment":"The notation 'epsilon_i ~ N(sigma_i, 0)' is nonstandard and misleading; it should be written as a zero-mean Gaussian with a given variance, e.g., N(0, sigma_i^2), and the trusted region in Eq. (8) should be defined consistently as a 1-sigma ellipsoid in the measurement-coordinate space.","section":"Sec. III-C, Definition 5"},{"comment":"The sign of the term inside the square root and the definitions of b_i are unclear; please check the algebra and ensure the constraint reformulation in Eqs. (13)-(14) is consistent with the original constraint in Eq. (12).","section":"Sec. III-C, Eqs. (13)-(14)"},{"comment":"The entry for 300 steps with the geometric method is reported as '0.97' without units; the table would be clearer if it stated that all times are in seconds and if the missing unit were added.","section":"Sec. IV-B, Table II"},{"comment":"There is a typo in 'con sider' which should be 'consider.'","section":"Sec. IV-C, wrench experiment paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is an early arXiv version and would benefit from closer editing. The main concern for the editor is that the title and abstract emphasize robustness to force uncertainty and the geometric framework, but the experiments do not separate the estimator's contribution from the robust planner's contribution; without an ablation the central theoretical claim is unsubstantiated. I see no circularity issue: the framework is not fit to the success data. The lack of reported sigma_i is a straightforward fix and should be required."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe short version: GeoDEx is a coherent estimation-planning-control framework that makes a real difference on hardware—wrench grasp success goes from 20% to 80%, cylinder from 0% to 60%, and cube pivoting stays under 5 degrees of error. That is not noise; the effect is large and consistent across tasks. But the paper's headline contribution, the uncertainty-aware robust planner, is never directly tested. The hardware comparisons always use the full system against raw measurements, and the noise variances (sigma_i) that define the ellipsoid are never reported. So we know the total stack works, and we don't know which part of it does the work.\n\nWhat's new: the combination of projecting force measurements onto the force-equilibrium plane and then using ellipsoidal uncertainty sets to plan forces that remain feasible is not in the prior tactile-manipulation literature. Individually the tools are standard—QM, QR decomposition, pyramid friction cones—but the packaging for tactile force control with extrinsic contacts is a legitimate extension, and the 14x speed-up over SOCP is a real practical point. The paper is also honest about its limitations (sensor saturation, unknown contact location noise, quasi-static assumption).\n\nWhere it gets soft:\n\n1. The Gaussian noise assumption (Definition 5) is at odds with the paper's own sensor characterization: activation-threshold jumps, hysteresis, partial-contact offsets. The authors say in the Discussion that modeling the error 'is extremely difficult, if not impossible.' That is fine as a caveat, but then the ellipsoid's shape and size are arbitrary, and the feasibility constraints in Problems 3 and 4 provide no tested guarantee against the actual error distribution. The sigma_i values are never given, so a reader cannot even check conservativeness.\n\n2. No ablation isolates the robust planner. The success could be entirely from the FE-plane projection in the estimator, which corrects the force balance. The stress-test note is right: without swapping the robust planner for a nominal one (keeping the estimator), the uncertainty-aware planning is not load-bearing evidence.\n\n3. The extrinsic estimation objective (Eq. 7) is underspecified—it minimizes a sum of distances but the derivation of the subspace basis and the weights is compressed to the point of being hard to follow.\n\nNone of these are fatal. The hardware results are credible, the writing is clear, and the authors acknowledge several of these gaps themselves. But a referee should ask for the sigma_i values, an ablation, and a cleaner derivation of Eq. 7.\n\nWho this is for: people working on tactile-based force control and dexterous manipulation with imperfect sensors. They will get useful ideas and a working recipe, even if the theoretical robustness guarantee is more aspirational than demonstrated.\n\nRecommendation: send it to review. It deserves referee time, but conditionally—major revision to address the missing ablation and unreported parameters.","headline":"A coherent geometric estimation-plus-planning stack that clearly helps on hardware (wrench grasp success 20% to 80%, cylinder 0% to 60%), but the uncertainty-aware planner is never isolated or validated, so the central robustness claim outruns the evidence.","tokens_in":14821,"tokens_out":2339,"would_cite":true,"duration_ms":22297,"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":"A geometric framework lets robots grasp and pivot objects using inaccurate tactile force readings alone.","keywords":["tactile sensing","dexterous manipulation","force control","extrinsic manipulation","force estimation","force equilibrium","admittance control","geometric planning"],"falsifier":"Measure each taxel's error distribution against a force-torque reference across a grid of contact locations and force magnitudes on the same fingertip hardware. If the empirical error cloud is not zero-mean and ellipsoidal, or if errors are correlated across taxels and contacts, then the trusted ellipsoid of Definition 5 will systematically misplace the true force and the success-rate improvement should degrade. A direct control test would inject a known constant force bias into one fingertip's readings and check whether the estimator still drives the object to equilibrium.","tokens_in":13877,"feed_emoji":"🤖","tokens_out":6651,"duration_ms":63966,"temperature":0.7,"pith_summary":"This paper argues that force-controlled dexterous manipulation does not require accurate tactile force sensors, only a correct geometric accounting of where the true contact forces must lie. GeoDEx represents all forces that can balance an object's weight as a plane, all possible normal-force readings as a cone, and the sensor's uncertain measurements as an ellipsoid, so that force estimation and planning become projections and linear feasibility problems on that plane. On hardware with noisy tactile fingertips, the framework raises wrench-grasp success from 20% to 80% and cylinder-grasp success from 0% to 60%, and keeps cube-pivot angular error under 5 degrees, while the same raw readings fail when fed directly to an admittance controller. The practical payoff is that precise force control becomes possible with cheap, noisy tactile sensors instead of embedded force-torque sensors.","feed_headline":"Grasp success leaps from 20% to 80% with noisy touch data","feed_subtitle":"A plane-cone-ellipsoid framework recovers true contact forces and plans safe grasps and table-assisted pivots.","key_machinery":"The load-bearing objects are the force-equilibrium plane (the hyperplane of contact-force vectors that balance gravity), the measurement cone (the set of readings consistent with normal-only force sensors at known contacts), and the trusted measurement ellipsoid (a Gaussian-variance ellipsoid in measurement space bounding where the true normal forces can lie). Propagating the ellipsoid onto the equilibrium plane converts force planning into a linear feasibility problem in FE-coordinates, and the same plane projection supplies the force estimate used by an admittance controller. The controller then moves each fingertip along the direction dictated by the error between desired and estimated forces, rather than the raw tactile readings.","core_discovery":"The central claim is that a force-equilibrium constraint turns unreliable normal-force readings into a solvable estimation and planning problem. Given contact positions and normals, every force vector that cancels gravity lies on a hyperplane, the force-equilibrium plane; projecting the observed tactile forces onto this plane yields force estimates that respect equilibrium even when individual sensor readings are biased. Planning is handled by enlarging each measurement into a trusted ellipsoid derived from assumed per-contact noise variance, projecting that ellipsoid onto the equilibrium plane, and finding a center whose entire projected ellipsoid stays inside the friction-cone and minimum-force constraints. Because the reformulated constraints are linear, the planner solves a linear feasibility problem rather than a general nonlinear or second-order cone program. The hardware experiments support the paper's assertion that this pipeline enables successful grasping and extrinsic pivoting where direct control from raw measurements fails.","pith_inferences":["The same geometric principle should transfer to other noisy or biased contact-sensor types, as long as contact locations and normals remain accurate enough to define the equilibrium plane and measurement cone.","A testable trade-off follows: engineering effort could shift from making tactile force readings accurate to making contact geometry accurate, since the paper identifies contact-location and normal uncertainty as its main unmodeled limitation.","The noise ellipsoid parameters could be learned or adapted online from force-torque ground truth during a calibration phase, letting the planner shrink its trusted region as the sensor's actual error distribution is identified.","A further extension would replace the Gaussian trust model with a distributionally robust set, which would address the threshold, hysteresis, and bias errors the paper characterizes but does not model."],"forward_implications":["A hand with normal-only tactile fingertips can perform quasi-static force-controlled grasps, so fingertip force-torque sensors are not necessary for these tasks.","Unknown extrinsic contact forces, such as object-table contact, can be folded into the same plane-cone-ellipsoid machinery, enabling table-assisted pivoting with the same estimator and planner.","Because the geometric reformulation turns planning into linear feasibility, it runs about 14 times faster than a direct second-order cone program, making online re-planning practical.","On hardware, using estimated forces instead of raw readings raises wrench-grasp success from 20% to 80% and cylinder-grasp success from 0% to 60%.","The framework tracks a cube pivoting about a table edge with RMS angular error under 1 degree in simulation and under 5 degrees on hardware."],"supporting_citations":[{"why":"Supplies the tactile fingertip sensor used in all hardware experiments, whose raw readings the framework must correct.","marker":"[13]"},{"why":"Documents a state-of-the-art tactile array whose force readings are inaccurate, motivating post-processing of imperfect readings.","marker":"[10]"},{"why":"Shows that even binary contact signals improve in-hand manipulation, supporting the use of salient but imperfect tactile information.","marker":"[2]"},{"why":"Demonstrates that discretized tactile force direction can drive in-hand reorientation, a related extraction of salient signal from noisy touch.","marker":"[3]"},{"why":"Provides a hardware extrinsic-manipulation result that requires a force-torque sensor and a single rigid object, the setup GeoDEx seeks to simplify.","marker":"[27]"},{"why":"Establishes that force information improves dexterous grasping under uncertainty, motivating force-aware planning.","marker":"[22]"},{"why":"Represents the learned force-estimation alternative that requires labeled force data, contrasted with the geometric post-processing approach here.","marker":"[18]"},{"why":"Supports the use of contact force information in model-based planning and control of in-hand manipulation.","marker":"[1]"}],"fun_headline_variants":["Geometry tames noisy touch: robots grasp and pivot reliably","Force uncertainty? Project onto a plane, plan with an ellipsoid","GeoDEx: 14x faster tactile manipulation despite noisy forces","Uncertain touch? This framework turns it into safe grasps","From unstable to successful: geometry beats raw tactile force"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All of the guarantees and the size of the trusted force region rest on the assumption that each contact's normal-force measurement error is independent and Gaussian with a known variance; the paper's own sensor characterization lists threshold jumps, hysteresis, and contact-dependent errors that are not Gaussian, and no hardware variance values are reported.","fun_headline_variants_meta":{"raw":{"variants":["Geometry tames noisy touch: robots grasp and pivot reliably","Force uncertainty? Project onto a plane, plan with an ellipsoid","GeoDEx: 14x faster tactile manipulation despite noisy forces","Uncertain touch? This framework turns it into safe grasps","From unstable to successful: geometry beats raw tactile force"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2128,"prompt_tokens":910,"completion_tokens":1218,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1132}},"tokens_in":526,"tokens_out":1218,"duration_ms":12133,"temperature":1.0,"reasoning_tokens":1132,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:36:49.036965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure each taxel's error distribution against a force-torque reference across a grid of contact locations and force magnitudes on the same fingertip hardware. If the empirical error cloud is not zero-mean and ellipsoidal, or if errors are correlated across taxels and contacts, then the trusted ellipsoid of Definition 5 will systematically misplace the true force and the success-rate improvement should degrade. A direct control test would inject a known constant force bias into one fingertip's readings and check whether the estimator still drives the object to equilibrium.","supporting_citations":[{"cited_title":"https://www.touchlab.io/","cited_arxiv_id":null,"evidence_quote":"Supplies the tactile fingertip sensor used in all hardware experiments, whose raw readings the framework must correct."},{"cited_title":"https://www.xelarobotics.com","cited_arxiv_id":null,"evidence_quote":"Documents a state-of-the-art tactile array whose force readings are inaccurate, motivating post-processing of imperfect readings."},{"cited_title":"Robot synesthesia: In-hand ma- nipulation with visuotactile sensing","cited_arxiv_id":null,"evidence_quote":"Shows that even binary contact signals improve in-hand manipulation, supporting the use of salient but imperfect tactile information."},{"cited_title":"General in- hand object rotation with vision and touch","cited_arxiv_id":null,"evidence_quote":"Demonstrates that discretized tactile force direction can drive in-hand reorientation, a related extraction of salient signal from noisy touch."},{"cited_title":"Robust learning of tactile force estimation through robot interac- tion","cited_arxiv_id":null,"evidence_quote":"Represents the learned force-estimation alternative that requires labeled force data, contrasted with the geometric post-processing approach here."},{"cited_title":"Hybrid learning-and model-based planning and control of in-hand manipulation","cited_arxiv_id":null,"evidence_quote":"Supports the use of contact force information in model-based planning and control of in-hand manipulation."}],"review_version":1}