{"id":"b9dc9c0e-f88d-4ace-8a1b-9240381f847c","arxiv_id":"2509.09484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A dual-arm vision-driven system bags objects into fabric bags by tracking and planning only the bag opening rim as a constant-perimeter ellipse, succeeding in 12 scenarios with the lowest misalignment in its comparisons.","lead":"A pair of robot arms learns to bag objects, such as boxes, cans, fruit, and bundled items, into fabric shopping bags by watching the bag's opening rim and adjusting on the fly. The system needs no model of the bag's material, only a camera, and plans the whole operation as a two-stage rim path, which could make automated bagging in packaging lines cheaper and more flexible.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unverified constant-perimeter assumption underpins all planned SOIs; if violated, rim-only representation is not sufficient.","rationale":"The paper's central claim is that the rim SOI suffices; this rests on the rim being a stable, near-elliptical, constant-perimeter curve. The reader flagged the same Sec. III assumption. I could not find a more load-bearing issue: the GMM extraction and Broyden Jacobian are standard and could be patched; the constant-perimeter assumption is baked into the optimization and every planned waypoint. A measurement test can settle it. Since the paper already received CONDITIONAL and the concern is testable rather than demonstrated false, I recommend keeping the verdict unchanged.","tokens_in":13896,"tokens_out":4638,"duration_ms":54581,"concrete_test":"Repeat Exp. 1 (coffee box) with the same setup; at each MPC step, before GMM downsampling, trace the bag rim in the dense point cloud (e.g., fit a closed polyline to the topmost contour) and compute its 3D length L_t. Also compute the GMM chord perimeter R_t. If max_t |L_t − R_0|/R_0 > 2% or max_t |R_t − R_0|/R_0 > 2% during pre-bagging and bagging, the constant-perimeter assumption is violated and the planned SOI may be infeasible; if both stay within 2%, the assumption holds and the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. III states 'the size of the bag opening remains relatively unchanged during the bagging process, staying close to its initial size.' This assumption is load-bearing: the bagging SOI is optimized against R (Eq. 7, ||L−R||^2), every regularized subgoal is forced to have nearly the same perimeter via C4 (Eq. 16, λ4=0.002), and the goal SOI is a translation of that same g† (Eq. 13). R itself is computed as a chord sum over the initial GMM rim, not as an arc length along the fabric. A sagging or partially occluded rim will produce a smaller observed chord perimeter even though the fabric edge length is constant; the MPC then tracks an ellipse whose chord perimeter cannot match the physical rim without stretching or folding. The paper reports no measurement of R_t during manipulation, so the assumed invariance is unverified. If it fails, the planned reference is a fictional shape and the adequacy of rim-only SOI—the central claim—is unsupported beyond the specific tested bags.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a complete dual-arm robotic system for bagging objects into fabric bags. The bag's opening rim is treated as a Structure-of-Interest (SOI); a GMM is used to estimate the rim from raw point clouds, a constrained optimization generates a 'bagging SOI' ellipse around the object's bottom, CBiRRT plans a trajectory from the initial rim to this ellipse and then to a translated goal SOI, and an MPC-based shape-servoing controller with an online Broyden-estimated deformation Jacobian tracks the planned subgoals. The central claim is that the rim-only SOI is an adequate representation of the fabric state for the bagging task. The system is evaluated in 12 scenarios with different objects and bag types, with success rates and comparisons to several baselines.","tokens_in":14155,"tokens_out":4622,"duration_ms":59202,"significance":"If the central claim is correct, the paper offers a practical and reasonably general solution to a challenging deformable-object manipulation task, without requiring a physics model or material parameters. The paper's strengths are its real-hardware validation across multiple objects and bag types, the large simulation study for SOI generation, the inclusion of baseline planners/controllers, and the availability of a video. The reported 8/8 manipulation success in most scenarios is encouraging. However, the load-bearing constant-perimeter assumption is stated but not verified, and the statistical basis for the comparative success-rate claims is weak. These issues do not invalidate the demonstrated system but leave the generality of the rim-only representation insufficiently supported.","major_comments":[{"comment":"The constant-perimeter assumption is load-bearing but unverified. Section III states that \"the size of the bag opening remains relatively unchanged during the bagging process.\" R is defined as a chord sum over the initial rim (Sec. V-A.2), not as fabric arc length. The bagging SOI optimization minimizes ||L-R||^2 (Eq. 7), every regularized subgoal enforces |R_x/R_y - 1| <= lambda_4 (Eq. 16), and the goal SOI is a translation of g-dagger (Eq. 13). If the physical rim sags, stretches, or is occluded, its observed chord perimeter changes even though the fabric length is constant; the MPC then tracks a planned ellipse that cannot match the physical rim. No online R_t measurement or adaptation is reported, so the central claim that rim-only SOI is adequate is not established beyond the specific bags tested. Please report R_t traces or add a re-estimation/re-planning mechanism.","section":"III, Eqs. (7), (13), (16)"},{"comment":"The headline success rates are quoted as x/8 or x/10 without confidence intervals or repeated trials. With n=8, a single failure changes the rate by 12.5 percentage points; observed differences such as 8/8 vs 7/8 are not statistically meaningful. Planning success rates are also based on a single set of 10 runs. This weakens the comparative claim of \"highest success rate.\" Please report repeated trials, confidence intervals, or a statistical test.","section":"VII-E, Tables I and II"},{"comment":"The text says that a uniform distribution is incorporated into the GMM to handle outliers, but Eq. (4) sums only over n_x Gaussian components and has no uniform term. Either the equation is missing a uniform component or the implementation description is inaccurate. Since SOI extraction is the perception backbone of the whole pipeline, this inconsistency needs clarification.","section":"IV, Eq. (4)"},{"comment":"The deformation Jacobian is assumed to have full column rank. This is not verified or enforced. For a slack bag, some arm motions may produce little or no rim displacement, making J_t rank-deficient; the Broyden update (denominator u_t^T u_t) also becomes ill-conditioned when the arm velocity command is small. Please provide safeguards, a rank check, or experimental evidence that rank deficiency does not occur in the reported scenarios.","section":"VI-B, Eq. (18)"}],"minor_comments":[{"comment":"The equation should read \"sum_j pi_j = 1\" rather than \"P pi_j = 1\".","section":"Eq. (4)"},{"comment":"The x/y/z axes are computed from three randomly selected vertices of the object bottom. A near-collinear selection makes the cross products in Eq. (5) ill-conditioned. A PCA-based or otherwise robust frame estimation would avoid this numerical fragility.","section":"V-A.1, Eq. (5)"},{"comment":"The note that g* is a simple translation of g-dagger and \"may involve a more complex configuration\" is important; please discuss this limitation in the conclusion as future work.","section":"VII-D"},{"comment":"R is used both for the bag perimeter (Sec. V-A.2) and for the MPC control-weight matrix (Eq. 19). Using distinct symbols would avoid confusion.","section":"Notation"},{"comment":"The flowchart is dense and the panel labels are small. Enlarging the figure or listing the steps in the caption would improve readability.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper describes a complete, well-engineered system and the experiments are encouraging for the demonstrated objects. The main technical concern is not the control/estimation math, which is standard, but the unverified constant-perimeter assumption on which all planned SOIs depend. I would like the revision to address this directly, either by measuring and reporting R_t over the manipulation or by adding an adaptive re-planning mechanism. The statistical weakness of the success-rate comparisons is secondary but should also be fixed. The paper's own limitations (solid-color detection, simple g* translation) should be acknowledged in the main text rather than left as asides."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, honest systems paper. The genuinely new thing is the full bagging formulation: using the bag rim (SOI) as the state, generating a bagging ellipse perimeter-matched to the initial rim and aligned with the object bottom, splitting the motion into pre-bagging and bagging stages, and tracking planned ellipses with an MPC shape servo. The building blocks are published, but nobody has put them together for fabric-bag bagging, and the video and the 12 experiments show the idea works on the tested objects.\n\nThe paper does several things well. The math is standard and internally consistent. They use Broyden to estimate the deformation Jacobian online, so the controller does not assume a material model. They measure the outcome externally (bag wraps object, alignment error), so there is no circularity. They also state limitations explicitly — solid-color bag detection issues, the need for further bag types — which is more than many such papers do.\n\nThe soft spots are proportionate. First, the load-bearing assumption that the bag rim perimeter stays constant is stated in Sec. III and used everywhere (Eq. 7, Eq. 16, Eq. 13), but never measured during manipulation. Worse, the perimeter is computed as a chord sum, not an arc length along the fabric, so a sagging or occluded rim can make the chord perimeter shrink even when the fabric length is constant. For the demonstrated bags the assumption held well enough, but the paper gives no evidence about when it breaks. That is the main reason I would not take the central claim as a general result.\n\nSecond, the experimental base is thin: manipulation success is x/8, no confidence intervals, the baselines are re-implemented with modifications, no code or data, and the solid-color rim detector is an unspecified neural net. These are real weaknesses, but they are the usual weaknesses of a robotics system paper, not load-bearing errors.\n\nWho is this for? People working on deformable-object manipulation and industrial packaging. It deserves a serious referee: the claim is new, the experiments support it for the demonstrated cases, and the limitations are well-scoped. I would accept it for review, and ask the authors to measure the rim perimeter during manipulation and report per-trial results with uncertainty.","headline":"A competent, honest dual-arm bagging system paper with a novel integrated formulation; the rim-only state and constant-perimeter assumptions are plausible for the tested bags but unverified as a general result.","tokens_in":14729,"tokens_out":2023,"would_cite":false,"duration_ms":24092,"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":"Dual-arm bagging succeeds by tracking only the opening rim as an ellipse.","keywords":["deformable object manipulation","bagging","structure-of-interest","dual-arm robots","shape servoing","model predictive control","Gaussian mixture models","constraint-aware planning"],"falsifier":"Measure the true rim perimeter with vision during a bagging trial on a stretchy or slack bag; if the perimeter deviates beyond the constraint tolerance, the planned elliptical reference will diverge from the physical rim and the MPC will track a fictional shape, breaking the near-perfect success. A second check: test a bag whose rim cannot be approximated by an ellipse (e.g., a deeply creased or sagging opening) and observe whether planning and control fail.","tokens_in":13735,"feed_emoji":"🛍️","tokens_out":3646,"duration_ms":42472,"temperature":0.7,"pith_summary":"The paper claims that to bag an object with a fabric bag, a robot does not need to estimate the entire bag state; tracking the bag's opening rim, called the Structure-of-Interest (SOI), is sufficient. The system extracts this rim from noisy point clouds, generates elliptical target rim shapes around the object's base, plans a sequence of elliptical subgoals, and controls dual arms with a model-predictive shape servoing loop. In experiments, this approach bags boxes, cans, fruit, prisms, and bundled objects with 8/8 manipulation success in most of the 12 reported scenarios and an average final alignment error of 2.10 cm, outperforming comparison methods. If correct, this offers a practical route to deformable-object manipulation that avoids heavy full-state modeling.","feed_headline":"Robots bag objects by tracking only the bag's opening rim","feed_subtitle":"A dual-arm system plans elliptical rim states and controls them in real time, scoring 8/8 in most bagging trials.","key_machinery":"The Structure-of-Interest (SOI): the set of points along the bag's opening rim, treated as the entire bag state. The pipeline regularizes every planned subgoal into a standard ellipse (via the Regularization function) and enforces that the ellipse's perimeter stays close to the bag's initial rim perimeter (constraint C4). The local controller is an MPC-based shape servoing loop that uses a Broyden-updated deformation Jacobian to convert rim shape errors into dual-arm velocity commands.","core_discovery":"The central claim is that the bag opening rim, represented as a set of 3D points called the SOI, is an adequate state representation for the whole bag during a bagging task. The system does not rely on prior knowledge of bag material properties. It estimates the rim in real time using a Gaussian Mixture Model over raw point clouds, then solves an optimization to place a bagging ellipse around the object's bottom, ensuring the ellipse's perimeter matches the measured rim perimeter and contains the object's base vertices. Planned subgoals are regularized into ellipses with a perimeter-consistency constraint, and an MPC-based shape servoing controller, which approximates the deformation Jacobia","pith_inferences":["The rim-as-ellipse abstraction likely extends to other enveloping tasks such as covering, wrapping, or placing bags over objects, not just the bagging scenario demonstrated here.","A natural testable extension is an adaptive perimeter estimator: if the system tracked rim perimeter changes in real time, it could handle stretchy or slack bags where the fixed-perimeter assumption breaks.","The success with bundled objects suggests the SOI representation may work for arbitrary bottom cross-sections as long as the rim can be approximated by an ellipse; non-elliptical rim models could broaden applicability further.","Because the system requires an initially open bag and a visible rim, the same feedback could be used to learn a bag-opening primitive, closing the gap to fully autonomous bagging from a closed bag."],"forward_implications":["Deformable bag manipulation can be achieved with modest perception: only the rim needs to be tracked, not the full fabric surface.","The two-phase strategy with an intermediate bagging SOI (g†) improves manipulation success compared to single-phase approaches.","Constraint-aware planning that keeps perimeter consistent makes the planned elliptical rim shapes physically plausible and collision-free.","MPC tracking of a time-varying deformation trajectory outperforms controllers that assume a static desired shape.","The method transfers to different bag colors and to bundled/tilted objects, with success limited mainly by rim detection quality on solid-color bags."],"fun_headline_variants":["Dual-arm robot bags objects by tracking rim alone","Adaptive robot bagging tracks only opening rim","Rim tracking lets dual-arm robot bag varied items","Robot bagging without knowing bag material","Real-time rim control enables robust robotic bagging"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The bag opening's perimeter stays roughly constant at its initially measured size throughout the entire bagging process, and every rim state can be faithfully approximated as an ellipse.","fun_headline_variants_meta":{"raw":{"variants":["Dual-arm robot bags objects by tracking rim alone","Adaptive robot bagging tracks only opening rim","Rim tracking lets dual-arm robot bag varied items","Robot bagging without knowing bag material","Real-time rim control enables robust robotic bagging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":978,"prompt_tokens":700,"completion_tokens":278,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":207}},"tokens_in":444,"tokens_out":278,"duration_ms":4013,"temperature":1.0,"reasoning_tokens":207,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:02:01.674640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the true rim perimeter with vision during a bagging trial on a stretchy or slack bag; if the perimeter deviates beyond the constraint tolerance, the planned elliptical reference will diverge from the physical rim and the MPC will track a fictional shape, breaking the near-perfect success. A second check: test a bag whose rim cannot be approximated by an ellipse (e.g., a deeply creased or sagging opening) and observe whether planning and control fail.","supporting_citations":[],"review_version":1}