{"id":"43b46686-5e93-492d-a3ed-2513f95b3adc","arxiv_id":"2505.13889","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A learned LSTM that predicts cable shape and tension is combined with a zonotope-based trajectory optimizer to enforce collision and tension safety in simulated wire harness assembly.","lead":"Robot manipulators moving cables and wires must avoid stretching them and hitting obstacles, but most planners only predict shape, not tension. This paper adds a learned tension predictor to a reachability-based safe planner and reports better simulated assembly results, but its certified safety relies on an unproven error bound.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3's 'certified error bound' is set to a finite test-set maximum, so constraint (12) does not certify safety over the planner's continuous trajectory family.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing concern: Assumption 3 is asserted but not established. The paper's certification language depends entirely on this assumption, and the implementation replaces a needed uniform bound with a finite-sample maximum. This is not an internal inconsistency in the experiments, but it is a failure of the paper's central claim. The simulation results may still show a useful empirical method, but they do not support 'certifiably safe' as stated. The reader's REJECT verdict is therefore appropriate; my stress-test pass does not move it. The concrete test would probe whether Assumption 3 actually holds on the planner's domain, which is the minimal check needed to rescue the certification claim.","tokens_in":8046,"tokens_out":3054,"duration_ms":32173,"concrete_test":"Run an adversarial search over the planner's domain: for the trained tension model, maximize ||f_true(k, theta) - f_pred(k, theta)|| over k in K and randomized engine poses theta using CMA-ES or a dense grid over a discretized K, with f_true from the same PyBullet simulator used for training. Compare the found maximum to the epsilon used in Eq. (12). If any sample has residual greater than epsilon, Assumption 3 is violated and the zero-overextension result is not a certificate; if no violation is found after extensive search, the practical concern is weakened but still not a mathematical certificate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that constraint (12) is 'certifiably safe' because Assumption 3 gives ||f_true - f_pred|| <= epsilon for all t_i, k. The support for this assumption is missing. The paper states: 'In our implementation, epsilon is estimated from the maximum residual observed on the test set' (Section II.C.4). A maximum over a finite test set is not an upper bound over the compact trajectory family K, over randomized engine poses, or over the closed-loop receding-horizon state distribution encountered at execution. It is at best an estimate of empirical error; the true supremum over the planner's domain can be larger. If the true tension error exceeds epsilon on any planned trajectory, then ||f_pred(t_i;k)|| + epsilon <= f_lim does not imply ||f_true(t_i;k)|| <= f_lim, so the safety guarantee fails exactly in the worst case the paper claims to handle. No Lipschitz bound, conformal guarantee, or reachability-based bound on the learned tension model is supplied, and no formal verification, code, or data accompanies the claim. The contact projection in Section II.A.2 also has no certified effect on tension error. Therefore, the zero-overextension result in Table I is an empirical observation conditional on a particular choice of epsilon, not a certificate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a motion planning and control framework for deformable linear object (DLO) manipulation that jointly predicts DLO shape and tension with an LSTM, then integrates these predictions into a real-time trajectory optimizer based on polynomial zonotopes and the ARMOUR reachability framework. The optimizer enforces joint, collision, and tension safety constraints; tension safety is handled through a constraint of the form ||f_pred|| + ε ≤ f_lim, where ε is called a certified error bound. The method is evaluated in a PyBullet wire-harness assembly task over 100 randomized trials, reporting 76 task successes, zero robot collisions, and zero DLO overextensions, compared to 54 successes/26 overextensions for ARMOUR and 38 successes/3 overextensions for Learning Where to Trust.","tokens_in":8276,"tokens_out":3723,"duration_ms":37891,"significance":"If the safety certification were valid, this would be an important bridge between learned deformable-object models and reachability-based planning. The empirical results are promising and the combination of learned tension prediction with an ARMOUR-style planner is a sensible heuristic. However, the paper provides no real verification or code, and the central term \"certifiable\" is not supported by the evidence presented; the zero-violation result in Table I is to a large extent a consequence of enforcing the proposed constraint rather than independent evidence of safety.","major_comments":[{"comment":"The central safety claim rests entirely on Assumption 3, which postulates a certified error bound ε satisfying ||f_true(t_i;k) - f(t_i;k)|| ≤ ε for all time steps and all trajectory parameters k. The implementation then sets ε to the maximum residual observed on the test set. A finite test-set maximum is not an upper bound over the compact trajectory family K, the randomized engine poses, or the closed-loop receding-horizon state distribution encountered at execution; without a Lipschitz bound, a conformal guarantee, or a reachability-based bound, Eq. (12) does not imply ||f_true|| ≤ f_lim. The paper therefore does not justify the word \"certifiably\" in the title, and Table I's zero-overextension outcome is conditional on a particular choice of ε rather than a certificate.","section":"II.C.4 (Assumption 3 and Eq. (12))"},{"comment":"The zero-overextension result is a direct consequence of constraining the planner with Eq. (12), so it cannot be used as evidence that the safety guarantee holds. To substantiate the certification claim, the authors should report the distribution of residual magnitudes on held-out and out-of-distribution task instances, the worst-case residual over the planner's trajectory family, the observed gap between ||f_hat|| + ε and f_lim, and ideally show a comparison against a baseline with a genuinely certified bound (e.g., conformal or Lipschitz-based). Without such evidence, the result is an empirical observation, not a verification of Assumption 3.","section":"Table I and Section III.C"},{"comment":"The contact handler projects penetrated nodes onto obstacle surfaces using a QP, but the manuscript does not specify whether tension is predicted before or after this projection. If tension is predicted from the original LSTM output, the projected shape changes are not accounted for in the error model; if tension is predicted after projection, the projection step itself becomes part of the learned pipeline and must be covered by the error bound. Either way, the missing specification leaves Assumption 3 without support for the contact-rich scenarios that are the focus of the paper.","section":"II.A.2 (Contact-aware modeling)"}],"minor_comments":[{"comment":"The notation N_t is used in Eq. (6) as if it were defined, but it is only later used in Eqs. (8)-(12); please define it explicitly as the set of time indices.","section":"Section II.C.4"},{"comment":"The variables n_p and n_t are both introduced in the cost function (Eq. 5) but their relationship is unclear; the sum runs to n_p while the time horizon is divided into n_t subintervals, which should be stated more clearly.","section":"Section II.C.3"},{"comment":"The column heading \"Failed w/o Violation\" should be written out as \"Failed without Violation\" for clarity.","section":"Table I"},{"comment":"Please specify how the wire harness tension is extracted from PyBullet (e.g., constraint force magnitude, elastic force at nodes), since the safety constraint depends on this value.","section":"Section III.A.3"},{"comment":"The norm used in Eq. (6) and Assumption 3 is not specified; because the constraint is on ||f||, define whether this is the Euclidean norm or a componentwise maximum, as the error bound may follow a different distribution for each.","section":"Section II.C.4"}],"recommendation":"reject","confidential_remarks":"The paper has merit as an empirical heuristic and the simulation results are encouraging, but the central \"certifiably safe\" claim is not established by the evidence. In my view the mismatch between the title/abstract and the actual guarantee is severe enough that the paper should not be published in its current form. A revised submission that either supplies a genuine certified bound (e.g., via conformal prediction or Lipschitz-based analysis) or honestly presents the method as a heuristic with careful empirical evaluation would be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline is that the paper's integration is plausible but the certification claim is not supported. The safety constraint Eq. (12) is conditioned on Assumption 3's epsilon, and the implementation sets epsilon to the maximum residual on the test set. That is a finite-sample statistic, not an upper bound over the planner's continuous trajectory family or over the randomized task distribution. Without a genuine bound (Lipschitz, conformal, verified), the zero-overextension results in Table I are an empirical consequence of the constraint being enforced with a chosen margin, not a certificate.\n\nWhat is new and worth credit: jointly predicting DLO shape and tension with an LSTM and feeding both into an ARMOUR-style polynomial zonotope optimizer is a sensible combination that I haven't seen in the two cited baselines. The contact-aware projection and the cost that includes DLO state are reasonable design choices. The simulation results are internally consistent and show the tension constraint reduces overextensions relative to ARMOUR, and the method is less conservative than Learning Where to Trust. That is a real, if incremental, empirical contribution.\n\nWhere it goes soft: the main flaw is above, and it is load-bearing. The paper would be fine if reframed as a heuristic safety-constrained planner with uncertainty quantification, but the title and abstract promise certification. Additional soft spots: the contact projection has no provable effect on tension error; the evaluation is simulation-only with no code or data; and the reference list includes some unrelated self-citations (the drying papers), which does not help credibility.\n\nBottom line: this deserves serious peer review, but with a clear directive to the authors to either supply a real certificate or strip the certification language. The empirical framework is salvageable and the problem is important. I would not cite it as a certified method, but I might cite it as an example of learning-based tension-aware planning if the framing is fixed.\n\nBring it to reading group? Maybe, as a case study in why test-set error is not a safety certificate.","headline":"A sensible learning-plus-reachability integration for DLO manipulation that overclaims 'certifiable' safety by setting the error bound to a test-set maximum.","tokens_in":8808,"tokens_out":3032,"would_cite":false,"duration_ms":28939,"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 robot can manipulate deformable linear objects safely by predicting both future shape and tension and rejecting any planned trajectory whose worst-case tension exceeds a safety limit.","keywords":["deformable linear objects","tension prediction","certifiably safe motion planning","polynomial zonotopes","reachability analysis","wire harness assembly","contact-rich manipulation","LSTM prediction"],"falsifier":"Run the planner in an environment with a force sensor on the fixed end of the DLO, on a set of task configurations and trajectory parameters not used in training; if any trajectory is found where the predicted tension satisfies $\\lVert f(t_i;k)\\rVert + \\varepsilon \\le f_{\\text{lim}}$ while the measured true tension exceeds $f_{\\text{lim}}$, then the certified safety claim is false as stated.","tokens_in":7774,"feed_emoji":"🤖","tokens_out":6337,"duration_ms":59206,"temperature":0.7,"pith_summary":"The paper claims that a robot can manipulate cables, ropes, and other deformable linear objects safely in cluttered, contact-rich scenes by predicting both the future shape and the internal tension of the object, not just its shape. It builds a planner that evaluates whole families of robot trajectories, using a learned model to forecast DLO shape and tension, and rejects any trajectory whose predicted tension could exceed a safety limit. The authors test this in a simulated wire-harness assembly task: over 100 randomized trials the robot reaches the goal 76 times, never collides, and never overstretches the harness. If the claim holds, tension prediction is the missing ingredient that lets robots exploit necessary contact with the environment without sacrificing safety.","feed_headline":"Cable-tension forecasts yield zero safety violations in 100 trials","feed_subtitle":"A robot arm planning with learned shape and tension predictions assembles wire harnesses with no collisions or overstretching.","key_machinery":"The central mechanism is the joint shape-and-tension predictor and its integration with polynomial zonotope reachability. Polynomial zonotopes are set representations that overapproximate a family of trajectories parameterized by $k$ as polynomial functions of time and parameters, which lets the optimizer propagate uncertainty through the robot dynamics and the DLO tension model. The learned predictor is an LSTM-based network with a quadratic-programming contact handler, trained on 2,000 simulated trajectories, that produces the state and tension forecasts used in the cost function and safety constraints. The load-bearing identity is the safety constraint $\\lVert f(t_i;k)\\rVert + \\varepsilon \\le f_{\\text{lim}}$, which turns a predicted tension bound into a worst-case guarantee against overstretching, and Assumption 3 is the premise that such an $\\varepsilon$ exists uniformly.","core_discovery":"The paper's central claim is that DLO manipulation can be made certifiably safe by adding tension prediction to shape prediction inside a reachability-based trajectory optimizer. The learned model takes the current DLO state and the planned end-effector path and outputs both the future DLO shape and the tension profile, with a contact handler that projects penetrated nodes back onto obstacle surfaces. The planner overapproximates the reachable sets of joint configurations, velocities, inputs, occupancy, and DLO tension using polynomial zonotopes over a continuum of trajectory parameters. Safety is enforced by constraints including joint limits, collision avoidance, and the new constraint $\\lVert f(t_i;k)\\rVert + \\varepsilon \\le f_{\\text{lim}}$, where $\\varepsilon$ is an assumed certified bound on tension-prediction error. Under this constraint, even a worst-case underestimate of tension stays below the overstretch limit. In 100 simulated trials of wire-harness assembly, the framework reported a 76% success rate with zero robot collisions and zero DLO overextensions.","pith_inferences":["A formal safety certificate would require replacing the finite-test-set $\\varepsilon$ with a guaranteed uniform error bound, for example via conformal prediction or verified neural-network bounds; the paper's implementation is best read as an empirical safety margin rather than a mathematical certificate.","The same joint shape-and-tension forecast could be used as an online monitor: a wrist force or torque sensor could check whether true tension tracks the predicted profile and trigger a re-plan if the discrepancy approaches $\\varepsilon$, helping close the sim-to-real gap.","Model accuracy and safety margin trade directly against task success: a larger $\\varepsilon$ shrinks the feasible trajectory set, so improving tension prediction through better contact modeling or more data should raise the success rate in tight environments."],"forward_implications":["Executing a planned trajectory guarantees, under Assumption 3, no joint-limit violations, no collisions, and no tension above the safety threshold.","Robots can use necessary contact with the environment during DLO manipulation instead of treating all contact as risky, because the tension forecast distinguishes harmful overstretching from acceptable contact.","The planner's tension constraint directly reduces the risk of damaging the DLO or the robot, which matters in safety-critical applications such as assembly and surgical suturing.","Allowing necessary contact while still enforcing safety can raise task success compared with conservative contact-avoiding planners.","The same optimization structure can be reused for a variety of robot morphologies and trajectory families, with only the learned prediction model retrained for a new DLO task."],"supporting_citations":[{"why":"Supplies the polynomial-zonotope reachability analysis and robust controller that the proposed method extends with DLO shape and tension predictions.","marker":"[25]"},{"why":"Provides a learning-based baseline that classifies risky motions; the paper compares its conservative behavior and lack of safety guarantees against the new framework.","marker":"[22]"},{"why":"Gives the standard manipulator dynamics equations used to model the robot with the DLO reaction force at the end-effector.","marker":"[24]"},{"why":"Motivates tension estimation by showing force information is needed to manipulate DLOs in fixture-rich environments.","marker":"[15]"},{"why":"Supports the premise that environmental contact is often necessary in DLO manipulation, justifying the contact-aware design.","marker":"[5]"},{"why":"Represents prior DLO manipulation by interleaving prediction, planning, and control that the proposed framework upgrades with reachability-based safety.","marker":"[21]"}],"fun_headline_variants":["Joint shape-tension planning achieves zero safety violations in 100 tests","Certified safety in cable manipulation via joint tension prediction","Reachability-based tension control erases safety violations in wire assembly","Zero collisions and overstretch: tension-aware planner in 100 trials","Zero safety violations with joint shape-tension planning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole safety certificate rests on Assumption 3, that the learned tension model's error is always below epsilon for every time step and every trajectory the planner could choose, but the paper sets epsilon to the largest error on a finite test set, which does not prove a bound over the continuous trajectory family.","fun_headline_variants_meta":{"raw":{"variants":["Joint shape-tension planning achieves zero safety violations in 100 tests","Certified safety in cable manipulation via joint tension prediction","Reachability-based tension control erases safety violations in wire assembly","Zero collisions and overstretch: tension-aware planner in 100 trials","Zero safety violations with joint shape-tension planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001029,"raw_usage":{"total_tokens":4327,"prompt_tokens":928,"completion_tokens":3399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":3314}},"tokens_in":544,"tokens_out":3399,"duration_ms":23146,"temperature":1.0,"reasoning_tokens":3314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:08:14.113528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the planner in an environment with a force sensor on the fixed end of the DLO, on a set of task configurations and trajectory parameters not used in training; if any trajectory is found where the predicted tension satisfies $\\lVert f(t_i;k)\\rVert + \\varepsilon \\le f_{\\text{lim}}$ while the measured true tension exceeds $f_{\\text{lim}}$, then the certified safety claim is false as stated.","supporting_citations":[{"cited_title":"Can't Touch This: Real-Time, Safe Motion Planning and Control for Manipulators Under Uncertainty","cited_arxiv_id":"2301.13308","evidence_quote":"Supplies the polynomial-zonotope reachability analysis and robust controller that the proposed method extends with DLO shape and tension predictions."},{"cited_title":"Learning where to trust unreliable models in an unstructured world for deformable object manipulation,","cited_arxiv_id":null,"evidence_quote":"Provides a learning-based baseline that classifies risky motions; the paper compares its conservative behavior and lack of safety guarantees against the new framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the standard manipulator dynamics equations used to model the robot with the DLO reaction force at the end-effector."},{"cited_title":"Feel the tension: Manipulation of deformable linear objects in environments with fixtures using force information,","cited_arxiv_id":null,"evidence_quote":"Motivates tension estimation by showing force information is needed to manipulate DLOs in fixture-rich environments."},{"cited_title":"Robotic manipulation planning for shaping deformable linear objects withenvironmental contacts,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that environmental contact is often necessary in DLO manipulation, justifying the contact-aware design."},{"cited_title":"Manipulating deformable objects by interleaving prediction, planning, and control,","cited_arxiv_id":null,"evidence_quote":"Represents prior DLO manipulation by interleaving prediction, planning, and control that the proposed framework upgrades with reachability-based safety."}],"review_version":1}