{"id":"7cce5017-9feb-40e5-a02e-0a6f41ecfc8f","arxiv_id":"2506.22176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tracking-based, demonstration-free robot system ties an overhand knot from unseen rope configurations with 50% success in 16 trials.","lead":"KnotDLO ties an overhand knot with one robot arm by tracking the rope's current shape and computing grasp and target waypoints from that shape's topology. It offers an interpretable, training-free alternative to learned rope manipulation, though its 50% success rate on 16 trials is below the 66% learning-based baseline it compares against.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The X move's dependence on topology is the weakest link: the paper admits tracking fails after RII and replaces the undertip computation with a polygon-centroid heuristic; this fallback is load-bearing for the reported 50% success rate.","rationale":"The reader's weakest assumption identifies precisely the same load-bearing concern: the planner assumes the tracked curve S accurately reflects DLO topology, especially at self-crossings, but Section III admits tracking fails after RII and substitutes a polygon-centroid heuristic for the undertip used in the X move. The per-move success rates show the X move is the most failure-prone (0.615), and the overall 0.500 is the product of the three move rates, so the heuristic directly conditions the headline result. The paper is honest about this limitation, but it means the central claim of a topological, interpretable system is not fully supported for the final move. A small sample size (16 trials) further weakens the statistical strength, but that is secondary; the tracking fallback is the more fundamental issue because it challenges the method's core premise. This concern does not overturn the reader's conditional verdict; it reinforces it. I propose an ablation that would directly test whether the centroid heuristic is necessary, which would settle whether the topology assumption is actually load-bearing in practice. No independent artifacts are provided, so such a test would require the authors' system or a reimplementation, but it is the clearest way to validate the central claim.","tokens_in":6615,"tokens_out":11212,"duration_ms":118245,"concrete_test":"Re-run knot tying on at least 16 new initial configurations under two conditions: (A) the X move uses the topology-derived undertip computed by Eq. (3) (tip closest to the crossing bottom), and (B) the X move uses the paper's polygon-centroid heuristic. If condition A's success rate is statistically indistinguishable from B's, the tracking assumption is not the bottleneck. If A succeeds substantially less or fails outright, the centroid heuristic is load-bearing, confirming that the reported success depends on a non-topological fallback and that the 'topological state tracking' claim does not extend to the decisive move.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that KnotDLO achieves 50% success using topological DLO state tracking from perception as input. The weakest point is the X move, which the paper itself says cannot rely on the tracked topology: after RII, tracking often fails to initialize the crossing topology due to depth-resolution sensitivity, so the undertip grasp point is computed from the centroid of a convex polygon of tracked points instead of from the topological relation. This is not a trivial implementation detail. The X move has the lowest per-move success rate (0.615), and the overall success rate is exactly the product of the per-move rates (0.937 × 0.867 × 0.615 = 0.500). If the topology estimate is wrong at this step, the waypoint plan breaks, and the system's claimed interpretable, topology-driven planning is weakened. The 50% success rate is therefore contingent on a non-topological fallback for the most failure-prone move, so the claim that topological tracking is the enabling input is conditional and needs empirical qualification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents KnotDLO, a one-handed overhand knot tying system that uses TrackDLO's real-time piecewise-linear DLO tracking to compute grasp and target waypoints from the current DLO shape, without learning from demonstrations. The planner composes three primitives—Reidemeister Move I, Reidemeister Move II, and the Cross Move—based on the tracked topology and geometry. In 16 trials from previously unseen symmetric curved configurations, the system tied an overhand knot in 8 trials (50% success), with per-move success rates of 0.937, 0.867, and 0.615. The authors claim this is the first knot tying method using topological DLO state tracking from a perception system as input, and they emphasize interpretability, repeatability, and occlusion robustness.","tokens_in":6812,"tokens_out":4532,"duration_ms":47575,"significance":"If the reported result is robust, the contribution is notable: a training-free, interpretable alternative to learned knotting policies, with an explicit mapping from tracked DLO topology to waypoint poses and an empirical evaluation rather than a fitted prediction loop. The 50% success rate is honestly positioned against a learned baseline (66% for DDOD), and the per-move breakdown is a useful diagnostic. The significance is tempered, however, by the small trial count, the absence of confidence intervals, the lack of a quantitative occlusion analysis, and the paper's own admission that the most failure-prone move (X) relies on a non-topological polygon-centroid fallback when tracking fails after RII. These issues do not invalidate the contribution, but they do mean the central claims as stated are stronger than the evidence currently supports.","major_comments":[{"comment":"The manuscript states that tracking often fails to accurately initialize the crossing topology after RII due to depth-resolution sensitivity, and therefore the undertip grasp point S(s_cb,tip) is computed as the nearest tracked node to the centroid of a convex polygon formed from the tracked points, rather than from the topological relation. Since the X move has the lowest per-move success rate (0.615), and the total success rate is exactly the product of the per-move rates (0.937 × 0.867 × 0.615 = 0.500), this fallback is load-bearing for the headline result. The paper should report how many of the 16 trials triggered the fallback, the success rate conditional on fallback versus topological initialization, and how the fallback affects the claim that topological DLO state tracking is the enabling input for the planner.","section":"Section III (X move description)"},{"comment":"The headline 50% success rate is based on 16 trials, but no confidence intervals or exact counts are reported. The implied counts are RI 15/16, RII 13/15, X 8/13, and total 8/16; the 95% confidence interval for 8/16 is roughly [0.28, 0.72], so the difference from the DDOD baseline of 0.66 is not statistically meaningful as reported. Please report exact trial counts for each move, binomial confidence intervals, and a statement of how failures were attributed to perception, planning, or execution, including whether a failed move prevented later moves from being attempted.","section":"Section IV, Table I"},{"comment":"The comparison with DDOD, GSP, and Imitation is presented as a table of overall success rates without stating the experimental conditions for each baseline. Knot-tying success depends strongly on the robot, DLO material and length, initial configuration distribution, camera setup, and evaluation protocol, so a bare table of rates does not support a competitive claim such as 'compared to the success rate of the learning-based state-of-the-art method of 66%.' Please specify the setup for each baseline, or explicitly reframe the table as an informal reference point rather than a controlled comparison.","section":"Section IV, Table II and Introduction"},{"comment":"The abstract claims the method is 'robust to occlusion' and 'repeatable for varying rope initial configurations,' but the experiments do not measure occlusion severity or systematically vary initial configurations beyond starting from symmetric curved shapes with one taped tip. The paper should define the range of tested variation (e.g., initial curvature, camera viewpoint, self-occlusion duration) and report the failure breakdown by cause. Without this, the abstract's robustness and repeatability claims exceed the evidence presented.","section":"Abstract and Section IV (claims of robustness and repeatability)"}],"minor_comments":[{"comment":"The typeset equation for the geodesic distance is difficult to parse; in particular, the summation in the second case is not clearly rendered. Please check the equation formatting so that the piecewise definition is unambiguous.","section":"Equation (3)"},{"comment":"The sentence 'Manipulation assumes the perceived topology of the DLO is accurate before performing RI and RII, but not before performing RIII' refers to RIII, but the executed sequence is RI, RII, and X; RIII is never performed. This should be corrected to refer to the X move or removed.","section":"Section IV, final paragraph"},{"comment":"The figure captions are too brief to identify the seven curvilinear lengths (s=0.5, s=λ, s=1−λ, s_cb, s_ct, s_cb,tip, s_ct,tip) on the depicted rope. Annotating a representative image would substantially improve interpretability, which is a central claimed advantage.","section":"Figures 2 and 5"},{"comment":"The hand-selected constants λ=0.1, γ=0.4, r=5, and M=30 appear only in the experiments section; a sentence explaining how these values were chosen, or a sensitivity study, would clarify the claim of independence from task parameterizations.","section":"Section IV, parameter values"},{"comment":"There is a typo in 'tying an an overhand knot'; also, the abbreviation DDOD in Table II is not defined at first use in the text.","section":"Section IV, text"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is concise and reads more like a workshop or conference contribution than a full journal article. The core idea is sound and the experimental honesty is appreciated, but the X-move fallback, the lack of confidence intervals, and the uncontrolled baseline comparison are all load-bearing for the paper's central claims and should be addressed in revision. If the target venue is a journal, the experimental section will also need more depth, including a proper occlusion analysis and a sensitivity analysis of the hand-selected parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"KnotDLO is a genuine integration: TrackDLO's real-time curve tracking feeds seven topological waypoints that drive Reidemeister I, II, and a cross-move to tie an overhand knot, with no demonstrations and no learned policy. The 50% success rate in 16 trials is honestly reported, including per-move rates of 0.937, 0.867, and 0.615.\n\nWhat's new is the coupling of a perception-based topological state estimate to classical knot moves. The 'first to use topological tracking from perception' claim is plausible for this specific combination, but it needs qualification: after RII the tracking often can't resolve the crossing topology, so the X move's undertip is found with a polygon-centroid fallback rather than from the topology. Since overall success is exactly the product of the per-move rates, that fallback is load-bearing. The abstract's 'robust to occlusion' is not supported by any occlusion-specific experiment, and the comparison with learning-based baselines (DDOD at 66%) is uncontrolled — different setups, different rope lengths, different initial conditions. With n=16, the 95% CI on 50% is roughly 25–75%, so the gap is not meaningful.\n\nCredit where due: the authors explicitly state the tracking limitation and don't pretend the X move is topology-driven. The RI and RII success rates suggest the topological waypoint idea has merit. But the evidence is thin: no code or artifacts, no sensitivity analysis for λ, γ, r, M, no occlusion experiments, and the Cross Move is introduced without a citation.\n\nThis deserves a serious referee, not a desk reject. A good review would push for a larger trial set, a controlled baseline, and a clear decomposition of which steps actually rely on topology versus geometry. I'd bring it to a reading group, but I wouldn't cite it as a primary result.\n\nRecommendation: engage with it, but make the claims match the evidence.","headline":"Genuine integration but the headline success leans on a geometric fallback at the weakest move; worth a careful review with trimmed claims.","tokens_in":7344,"tokens_out":5361,"would_cite":false,"duration_ms":50893,"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":"Training-free robot ties overhand knots at 50 percent","keywords":["knot tying","deformable linear objects","topological waypoints","Reidemeister moves","visual tracking under occlusion","interpretable robot manipulation","training-free planning"],"falsifier":"Use a second overhead camera to record the true self-crossing order during the transition from the second move to the final poke-through move, and correlate each trial's final-move outcome with whether the tracked curve had the correct crossing topology; the central claim predicts that final moves succeed only in trials where the tracked topology was correct.","tokens_in":6433,"feed_emoji":"🪢","tokens_out":8718,"duration_ms":88122,"temperature":0.7,"pith_summary":"KnotDLO sets out to show that robotic knot tying does not need demonstrations, training, or a learned dynamics model: a planner reads the current tracked shape of a rope, converts that shape into a small set of grasp and target waypoints, and moves the rope through the topology moves that form an overhand knot. The rope is tracked as a piecewise linear curve, and every waypoint is derived from the curve's geometry and topology, so the policy is interpretable and does not depend on rope length or the exact starting configuration. In 16 trials from previously unseen configurations, the system tied the knot 8 times, a 50% success rate, with per-move success rates of 93.7%, 86.7%, and 61.5% for the three moves. The authors state this is the first knot-tying method to use topological deformable-linear-object state tracking from a perception system as its input. If the claim stands, a useful but imperfect knot-tying capability can be assembled from perception plus geometry, with failures traceable to a specific move rather than hidden inside a learned model.","feed_headline":"Training-free robot ties overhand knots at 50 percent","feed_subtitle":"It plans from the rope's tracked shape alone, using three topology moves and no demonstrations.","key_machinery":"The central object is the tracked piecewise linear curve $S$, a rope-shaped curve built from $M$ control points and parameterized by curvilinear length, together with a geodesic distance $\\rho(s_i,s_j)$ that measures separation along the rope rather than through 3D space. This curve is the single source of truth for planning: grasp positions are curve points, grasp orientations come from the curve's unit tangent under the semi-planar constraint, and target positions are computed from geometric combinations of selected curve points. The movement primitives are the Reidemeister moves from knot theory (twist, slide-over-loop, slide-over-crossing) plus a Cross Move for open ropes, and the planner selects seven special curvilinear lengths -- the midpoint, the two symmetric quarter points, the crossing bottom and top, and the under-tip and over-tip -- to turn each primitive into concrete waypoints. The mechanism does its work by decoupling visual reasoning from control: whatever the tracking system estimates about the rope's shape and self-crossings is converted deterministically into robot motion, which is why the authors call the result interpretable.","core_discovery":"The central claim is that an overhand knot can be tied by composing three movement primitives -- Reidemeister Move I, which twists the rope to add one crossing; Reidemeister Move II, which slides one strand over a loop to add two crossings; and the Cross Move, which pokes a rope tip through a loop to finish the knot -- and that the grasp and target poses for each primitive can be computed directly from the currently tracked piecewise linear curve $S(s_t;L)$ with no training and no human demonstrations. The planner selects seven curvilinear lengths on the curve, including the middle point, the two symmetric quarter points, and the crossing bottom and top, and turns them into waypoints using the local tangent direction under a semi-planar manipulation constraint. The authors report that this recipe tied an overhand knot in 8 of 16 trials from previously unseen symmetric curved starting configurations, and they identify the final Cross Move as the fragile step. They also state that KnotDLO is the first knot-tying system that takes topological DLO state tracking from a perception system as its input, which is what makes the waypoint plan interpretable rather than learned.","pith_inferences":["Beyond the reported trials, the total success rate 0.500 is nearly the product of the per-move rates ($0.937 \\times 0.867 \\times 0.615 \\approx 0.500$), consistent with move failures being roughly independent; if that independence holds, raising the Cross Move alone to 90% success would lift overall success to about 73%.","A test the paper does not run: vary depth noise or occlusion level while using a second camera to record the true self-crossing order, separating perception errors from execution errors and directly testing whether tracking accuracy is the load-bearing factor.","The same geometric recipe -- choose a knot type, choose a sequence of Reidemeister moves and Cross Moves, compute all waypoints from the tracked curve -- should extend to other knots such as the figure-eight, since the primitives are not specific to the overhand knot."],"forward_implications":["Because waypoints are computed from the tracked curve rather than learned, the same planner transfers to new rope lengths and new initial symmetric configurations without retraining.","The reported per-move success rates put the bottleneck on the final Cross Move (61.5%), so improving crossing-bottom localization and under-tip selection should raise overall success more than improving the earlier moves.","Decoupling visual reasoning from control means improvements in occlusion-robust rope tracking can be adopted without changing the planner.","The interpretable waypoint sequence can be used to collect labeled demonstration data for learning-based manipulation policies, which the authors propose as future work."],"supporting_citations":[{"why":"Supplies the tracked piecewise linear curve and self-crossing estimates from which every grasp and target waypoint is computed.","marker":"[24]"},{"why":"Provides the learning-based dense-descriptor baseline with 66% success that the 50% result is compared against.","marker":"[15]"},{"why":"Defines the Reidemeister moves that the planner composes into the overhand knot.","marker":"[25]"},{"why":"Provides the convex polygon centroid used to choose the under-tip grasp when the tracked crossing topology is judged unreliable.","marker":"[26]"}],"fun_headline_variants":["Robot ties overhand knots with three topology moves, no training","Shape-only planner succeeds at knot tying in 8 of 16 unseen tests","No demos, no RL: robot hits 50% on overhand knots from shape","Interpretable robot knot-tying: three moves, 50% success, no data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The planner assumes the tracked rope shape tells the truth about which strand is on top at the crossing the final move must poke through, and the paper itself notes that this estimate often fails after the second move because depth resolution is too coarse.","fun_headline_variants_meta":{"raw":{"variants":["Robot ties overhand knots with three topology moves, no training","Shape-only planner succeeds at knot tying in 8 of 16 unseen tests","No demos, no RL: robot hits 50% on overhand knots from shape","Interpretable robot knot-tying: three moves, 50% success, no data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001314,"raw_usage":{"total_tokens":5333,"prompt_tokens":905,"completion_tokens":4428,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":4342}},"tokens_in":521,"tokens_out":4428,"duration_ms":30041,"temperature":1.0,"reasoning_tokens":4342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:09:02.124123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a second overhead camera to record the true self-crossing order during the transition from the second move to the final poke-through move, and correlate each trial's final-move outcome with whether the tracked curve had the correct crossing topology; the central claim predicts that final moves succeed only in trials where the tracked topology was correct.","supporting_citations":[{"cited_title":"TrackDLO: Tracking Deformable Linear Objects Under Occlusion With Motion Coherence,","cited_arxiv_id":null,"evidence_quote":"Supplies the tracked piecewise linear curve and self-crossing estimates from which every grasp and target waypoint is computed."},{"cited_title":"Learning Rope Manipulation Policies Using Dense Object Descriptors Trained on Synthetic Depth Data,","cited_arxiv_id":null,"evidence_quote":"Provides the learning-based dense-descriptor baseline with 66% success that the 50% result is compared against."},{"cited_title":"Reidemeister, Knot Theory","cited_arxiv_id":null,"evidence_quote":"Defines the Reidemeister moves that the planner composes into the overhand knot."},{"cited_title":"Shapely,","cited_arxiv_id":null,"evidence_quote":"Provides the convex polygon centroid used to choose the under-tip grasp when the tracked crossing topology is judged unreliable."}],"review_version":1}