{"id":"ecb50e86-998b-4d2c-a508-77e1cf7055ad","arxiv_id":"2508.07319","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid force-position controller with a GAT-based learned dynamics model achieves high success rates in DLO shape control, but the force-space planning advantage is not isolated from waypoint decomposition.","lead":"This paper proposes a robot control method for reshaping deformable wires and cables by combining a force-space trajectory planner with model predictive control built on a learned graph-based dynamics model. The authors report success in simulation and on a real robot, but the key comparison to a position-space planning baseline is missing, so the force-space advantage is not isolated.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing position-space waypoint baseline confounds the central claim: force-space planning is never compared against simply interpolating keypoints in position space, so the observed success cannot yet be attributed to the hybrid force-position strategy.","rationale":"The reader's weakest_assumption concerns the physical feasibility of force-space interpolation and injectivity of the force representation. I think the more decisive issue is causal: even if force interpolation is feasible, the current baselines cannot attribute success to force-space planning. The missing position-space waypoint baseline is the cheapest decisive experiment and should be part of any conditional acceptance. This is partial agreement because the reader's rationale does mention the missing position-space waypoint baseline, but their stated weakest_assumption is about force-space interpolation. I would keep the CONDITIONAL verdict: the paper's reported results are promising, but this missing control prevents accepting the force-space mechanism as the explanation for the success rates.","tokens_in":10151,"tokens_out":6479,"duration_ms":70615,"concrete_test":"Implement PW-MPC: same EA-PE-GAT dynamics model and Algorithm 1, but generate 4 waypoints by linear interpolation between the current keypoint positions and target keypoint positions in R^{2m}, using the same waypoint-switching threshold (0.03 m), time-step limits, and controller gains as Section V-C3. Run all 10 simulation tasks for at least 10 random seeds and report per-task outcomes. If PW-MPC reaches 100% success with comparable RMSE, the force-space planning is unnecessary; if it fails on the same large-deformation tasks, the force representation is doing real work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the hybrid force-position strategy explains the improved success rate (Table I: 100% vs 12.5%/25%) is confounded by a missing ablation. Section V-C3 compares Hybrid only against P-MPC, which uses the same MPC but no intermediate waypoints, and against RL. The paper attributes success to planning in force space and converting waypoints to position space, but this mechanism is not isolated: any waypoint decomposition, e.g., linear interpolation in the 2m-dimensional keypoint position space, would also break a large deformation into smaller subgoals and could avoid the local optima that stall P-MPC. If a position-space waypoint MPC with the same dynamics model and hyperparameters succeeds at the same tasks, then force-space planning is not the cause of the results; the claim would reduce to waypoints help, which is not the paper's novelty. No error bars or repeated-task statistics are reported, and the large-deformation success rates are based on only 8 tasks, so this is not a minor statistical gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid force-position strategy for shape control of deformable linear objects (DLOs). It combines state trajectory planning in a 3D force space (end wrench) with MPC in position space, using learned MLP transformers (P2FT/F2PT) to convert between representations and a GAT-based dynamics model (EA-PE-GAT) for prediction. The authors report simulation results showing that the hybrid strategy achieves 100% success on eight large-deformation tasks, while position-only MPC and RL achieve only 12.5% and 25%, respectively, and they validate the approach on two real DLOs. The dynamics model is also compared against MLP and GAT baselines on multi-step prediction.","tokens_in":10504,"tokens_out":3290,"duration_ms":33087,"significance":"If the claims hold, the paper makes a useful contribution: a multimodal DLO state representation, an improved learned graph dynamics model, and a practical waypoint-based MPC scheme that outperforms position-only MPC and RL on large deformations. The dynamics-model comparison is informative and the availability of code/videos is a plus. However, the central claim that planning in force space is the cause of the improved success is not yet isolated by the experiments, and the statistical support is thin. The paper is likely of interest to the DLO manipulation community, but the main comparison needs strengthening before the claim is fully supported.","major_comments":[{"comment":"The central claim that force-space planning causes the success improvement is confounded. Hybrid is compared only with P-MPC, which uses the same MPC but no waypoints, and with RL. Any waypoint decomposition—e.g., linear interpolation in 2m-dimensional keypoint position space—would also break a large deformation into smaller subgoals and could avoid the local optima that stall P-MPC. Without a position-space waypoint MPC using the same dynamics model and hyperparameters, the results only show that adding waypoints helps; they do not isolate the force-space mechanism. Please add this ablation.","section":"§V-C3, Table I"},{"comment":"The statistical basis is very thin. Large-deformation success rates are computed from only 8 tasks (12.5% = 1/8, 25% = 2/8, 100% = 8/8), and small-deformation from only 2 tasks. No repeated trials, confidence intervals, or error bars are reported. The RMSE values in Table I (189 mm, 94.6 mm, 8.44 mm) are single numbers without variance. The contrast between 12.5% and 100% is a difference of one successful trial in eight. Please report per-task results, confidence intervals, or more tasks.","section":"§V-C3"},{"comment":"The force-space planner assumes that P2FT and F2PT are accurate and consistent inverses, and that the force representation is sufficiently injective for the DLO. Yet the F2PT average prediction error is 9.72 mm, while the success threshold in §V-C3 is 0.01 m (10 mm). Thus the converted waypoints are, on average, almost at the success threshold and may not correspond to reachable intermediate shapes. The paper does not verify that different DLO shapes with the same end wrench are absent, nor that linear interpolation in force space maps to feasible shapes. Please add an explicit consistency/injectivity check and report the error distribution.","section":"§V-C1 and §IV-C1"},{"comment":"The real-world validation is qualitative: four large-deformation experiments per object are shown, with no quantitative errors or comparison to baselines. The statement that models 'demonstrate consistent performance' is not supported by reported numbers. Please provide quantitative real-world errors or clearly label this as a demonstration.","section":"§V-C3 and §V-D"}],"minor_comments":[{"comment":"Typographical issues: 'a serious of keypoints' should be 'a series of keypoints'; 'represent by' should be 'represented by'; 'F orce Representation' has a spacing error.","section":"§III"},{"comment":"The simulation environment reports 'bending stiffness of 107' without units; please specify units (likely N·m² or similar).","section":"§V-A"},{"comment":"The real-world force sensor is described as attached to the UR5 flange to measure the fixed-end constraint force indirectly, while Fig. 2 shows the fixed end attached to a force sensor. Please clarify the measurement setup.","section":"§V-A and §V-D"},{"comment":"The paper says the dataset is split 8:1:1 but does not state whether the 3000 simulation trajectories are split before or after shuffling; please clarify to avoid temporally correlated splits.","section":"§V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid core and the dynamics-model evaluation is meaningful. The main obstacle to acceptance is the missing position-space waypoint baseline, which is essential to the paper's central claim. The small sample size compounds this. I would be willing to reassess after the authors add the ablation and stronger statistics; the structural issues seem addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is legitimately new work — planning intermediate waypoints in 3-D end-wrench space, converting them to position targets, and feeding them to an MPC whose dynamics model is a GAT with an explicit action encoder and a GRU property extractor. The dynamics ablation is the strongest part. EA-PE-GAT beats MLP, GA-Net, and two ablations consistently over multi-step horizons, and the action-encoder and property-extractor contributions are each supported by the comparisons. The P2FT/F2PT transformers are small MLPs with reported errors, and the authors even tried fancier transformer variants and reported those numbers too. Codes and videos are promised. That is real work, and the real-world validation on two DLO types is a plus.\n\nThe soft spot is exactly the one your stress test flags. Section V-C3 compares Hybrid to P-MPC (same MPC, no waypoints) and RL. That leaves the central claim — force-space planning is why large deformations succeed — not isolated. Any waypoint decomposition, including linear interpolation in keypoint position space, would break a large task into easier subgoals. If that baseline succeeds, the contribution shrinks to “waypoints help,” and force-space planning is an implementation detail. The success table is also thin: 8 large tasks, one run, no error bars. 100% vs 12.5% is striking, but on 8 tasks the uncertainty is large. The paper also says EA-PE-GAT outperforms “all other existing models” when it actually compared against four baselines, one of which is the relevant published GAT; that overclaim should be trimmed.\n\nA smaller modeling concern: the planner linearly interpolates in force space and relies on F2PT to turn each force waypoint into a reachable position waypoint. The transformers are accurate on average, but injectivity of the 3-D wrench representation for the DLO shape is not discussed. If two distinct configurations share the same end wrench, force-space planning can generate ambiguous waypoints. This is addressable and not fatal to the empirical pipeline.\n\nVerdict: worth a serious referee. The dynamics modeling contribution stands on its own, and the hybrid pipeline is a plausible advance; it just needs an ablation that changes the planning space while holding everything else fixed, plus repeated-trial statistics. The paper should be revised, not rejected on sight.","headline":"A useful hybrid force-position MPC for DLO shape control with a genuinely strong learned dynamics model, but the headline claim lacks a position-space waypoint ablation and the statistics are thin.","tokens_in":10866,"tokens_out":1651,"would_cite":true,"duration_ms":16637,"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":"The paper claims that planning a DLO's motion in a three-dimensional force space, then executing via MPC with a learned graph-attention dynamics model, makes large-deformation cable shaping reliable where position-only planning fails.","keywords":["deformable linear objects","shape control","force representation","model predictive control","graph attention networks","dynamics model learning","quasi-static manipulation","trajectory planning"],"falsifier":"Take two shapes that a robot can demonstrably reach but whose fixed-end wrenches are nearly identical (or identical), and ask the planner to move between them. If the force-space interpolation produces position waypoints that the MPC cannot track to the switching threshold, while direct position-mode MPC succeeds on the same pair, then the force representation is not carrying the information the strategy needs. A second, quantitative check: replace the interpolated force waypoints with the same number of randomly sampled wrenches of similar magnitude; if random waypoints succeed equally often,","tokens_in":10128,"feed_emoji":"🤖","tokens_out":6516,"duration_ms":60242,"temperature":0.7,"pith_summary":"This paper tries to establish a way to make robotic shape control of flexible wires and cables reliable even through large deformations, where direct position-based planning gets stuck. Its central proposal is to plan the motion in a low-dimensional force space: the two forces and one moment measured at the fixed end of the object are interpolated between the current and target shape, and the intermediate force values are converted back into keypoint positions that serve as sub-goals. An MPC controller then tracks those sub-goals using a learned dynamics model built from an explicit action encoder, a recurrent property extractor, and a graph attention processor. If the claims hold, the combination completes all tested large-deformation tasks in simulation with far lower shape error than position-only MPC or reinforcement learning, and the dynamics model predicts future shapes more accurately than the compared graph baselines.","feed_headline":"Hybrid force-position control hits 100% on large cable deformations","feed_subtitle":"Planning paths in force space plus graph-dynamics MPC finishes tasks where position-only methods stall.","key_machinery":"The hybrid force-position representation. The DLO's shape is a graph of keypoint nodes and distance-based edges; its force state is the wrench $W_e = [F_x, F_y, M_z]$ at the fixed end. Two MLP transformers, P2FT and F2PT, convert between graphs and wrenches. The planner linearly interpolates in wrench space, and the MPC uses the learned EA-PE-GAT dynamics model, whose explicit action encoder updates the grasped node's neighbors by a rigid-body transform, whose GRU-based property extractor infers node properties from history, and whose graph processor aggregates local interactions with multi-head attention. This pair of representations is what lets a large deformation be split into small, loc","core_discovery":"The central claim is that a deformable linear object's equilibrium shape is adequately summarized by the three-component wrench at its fixed end, so a straight-line path in that force space corresponds to a sequence of physically reachable shapes, even when the equivalent path in position space crosses local optima. The authors implement this with two learned MLP transformers: P2FT maps keypoint positions to the wrench, and F2PT maps a wrench back to keypoint positions. A trajectory planner linearly interpolates between the initial and target wrenches, and the MPC controller follows the resulting position waypoints. The controller uses the EA-PE-GAT dynamics model, which encodes the robot ac","pith_inferences":["If the fixed-end wrench is nearly a complete coordinate for quasi-static elastic-rod equilibria, then this approach is effectively path-planning on a low-dimensional equilibrium manifold; that predicts performance should degrade gracefully as tasks move toward 3-D or buckling regimes where the wrench is no longer injective.","The force-space interpolation is only as good as the P2FT/F2PT inverse consistency; a stricter test would be to close the loop by checking that converting a waypoint to positions and back yields the same wrench, and to use that residual as a waypoint-quality filter.","The small-deformation 'detour' suggests that the benefit of force-space planning is concentrated in large-deformation regimes, so a practical system would switch between force-mode and position-mode control based on the wrench distance or predicted shape error.","Because the dynamics model explicitly embeds the robot action as a rigid motion of the grasped neighborhood, the same architecture may transfer to other underactuated soft objects, such as cloth or rope in 3-D, with minimal changes."],"forward_implications":["Large-deformation DLO shape control can be decomposed into small-deformation subproblems by planning in force space, avoiding the local optima that stall position-only MPC and RL.","The EA-PE-GAT dynamics model, with explicit action encoding and a property extractor, yields multi-step prediction errors roughly an order of magnitude lower than MLP and visibly lower than GAT baselines, making it useful for model-based control.","The learned P2FT/F2PT transformers are fast enough (under half a millisecond) for real-time control, and simple MLPs outperform more complex transformer-based alternatives for this low-dimensional conversion.","A single framework with the same trained models transfers from simulation to real metal-strip and cable shaping, completing all trials without abrupt shape changes.","Small-deformation tasks can run slower than direct position-mode control because force-space planning can introduce unnecessary detours, pointing to adaptive waypoint selection as the natural next step."],"supporting_citations":[{"why":"Supplies the Jacobian-based global model baseline whose limitation to small deformations motivates the need for a force-space planner.","marker":"[3]"},{"why":"Provides the quasi-static Kirchhoff-rod equilibrium analysis that grounds the choice of the fixed-end wrench as the DLO's force state.","marker":"[6]"},{"why":"Establishes the graph-based keypoint representation of DLO state that the position representation builds on.","marker":"[5]"},{"why":"Provides the GA-Net baseline for dynamics prediction and the data-collection scheme the authors adopt for training.","marker":"[14]"},{"why":"Provides the real-world experimental setup with keypoint detection and external-contact graph dynamics that this framework extends.","marker":"[18]"},{"why":"Provides the MuJoCo physics simulation used for data collection and evaluation.","marker":"[21]"}],"fun_headline_variants":["Force-space planning beats position-only for cable shaping","Wrench-space path planning avoids cable shape dead ends","Hybrid force-position control with graph attention dynamics","MPC with learned graph model tames deformable cables","Cable shape control: plan in force space, follow with MPC"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Linear interpolation between the initial and target fixed-end wrenches yields a sequence of force waypoints that, when mapped back to positions, are shapes the robot can actually reach; this requires the learned force-to-position transformer to act as a reliable inverse and the wrench to identify a shape almost uniquely.","fun_headline_variants_meta":{"raw":{"variants":["Force-space planning beats position-only for cable shaping","Wrench-space path planning avoids cable shape dead ends","Hybrid force-position control with graph attention dynamics","MPC with learned graph model tames deformable cables","Cable shape control: plan in force space, follow with MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1688,"prompt_tokens":690,"completion_tokens":998,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":920}},"tokens_in":434,"tokens_out":998,"duration_ms":9426,"temperature":1.0,"reasoning_tokens":920,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:12:11.222141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two shapes that a robot can demonstrably reach but whose fixed-end wrenches are nearly identical (or identical), and ask the planner to move between them. If the force-space interpolation produces position waypoints that the MPC cannot track to the switching threshold, while direct position-mode MPC succeeds on the same pair, then the force representation is not carrying the information the strategy needs. A second, quantitative check: replace the interpolated force waypoints with the same number of randomly sampled wrenches of similar magnitude; if random waypoints succeed equally often,","supporting_citations":[{"cited_title":"Global model learning for large deformation control of elastic deformable linear objects: An efficient and adaptive approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the Jacobian-based global model baseline whose limitation to small deformations motivates the need for a force-space planner."},{"cited_title":"Quasi-static manipulation of a kirchhoff elastic rod based on a geometric analysis of equilibrium configura- tions,","cited_arxiv_id":null,"evidence_quote":"Provides the quasi-static Kirchhoff-rod equilibrium analysis that grounds the choice of the fixed-end wrench as the DLO's force state."},{"cited_title":"Offline-online learning of deformation model for cable manipulation with graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the graph-based keypoint representation of DLO state that the position representation builds on."},{"cited_title":"Learning graph dynamics with interaction effects propagation for deformable linear objects shape control,","cited_arxiv_id":null,"evidence_quote":"Provides the GA-Net baseline for dynamics prediction and the data-collection scheme the authors adopt for training."},{"cited_title":"Learning graph dynamics with external contact for deformable linear objects shape control,","cited_arxiv_id":null,"evidence_quote":"Provides the real-world experimental setup with keypoint detection and external-contact graph dynamics that this framework extends."},{"cited_title":"Mujoco: A physics engine for model-based control,","cited_arxiv_id":null,"evidence_quote":"Provides the MuJoCo physics simulation used for data collection and evaluation."}],"review_version":1}