{"id":"834dc119-b0d5-4ea8-b8c2-c02e2dd7fab2","arxiv_id":"1908.08509","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A Hessian-free correction to Rimon-Koditschek gradient dynamics guarantees convergence to the goal and obstacle avoidance for quadratic potentials with ellipsoidal obstacles of any eccentricity.","lead":"This paper modifies the standard potential-field navigation rule so a robot can reach its goal even when the obstacles are very flat ellipsoids, a case where the classical Rimon-Koditschek method can get stuck. The new rule only needs distances and relative positions, not curvature estimates, and the authors prove convergence for almost all starting points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3's proof is built on an explicit conjecture and does not establish the claimed unique unstable critical point on the repulsion-zone boundary; the Section IV-B cycle argument is likewise asserted, so Theorem 1 is not proved.","rationale":"The reader's weakest-assumption analysis is accurate and aligns with the most serious gap. Theorem 1's proof is a chain: global Lyapunov decrease outside V_k, local escape from epsilon-balls via Lemmas 1-3, and finite visits via Lemma 4 and the cycle argument. The weakest link is Lemma 3, which is the only component establishing that the repulsion-zone boundary does not contain an attracting set. Its proof explicitly relies on a conjecture about the location of a saddle (Appendix C), and the subsequent uniqueness argument is not a proof: it only rules out other critical points in a small gamma-neighborhood of the conjectured point and asserts the rest. The finite-visit argument in Section IV-B is similarly unsupported. These are not cosmetic; if ∂B_i^k contains a stable critical point or a closed orbit, the agent may fail to converge or even cycle indefinitely, directly contradicting Theorem 1. The numerical results in Section V are encouraging and show the dynamics often work, but they cannot establish the existence of a universal K for all configurations. The abstract's weaker 'almost every initial position' claim also does not rescue the proof because the same lemmas are needed. Therefore the appropriate verdict remains REJECT, with the possibility of ACCEPT if the proof of Lemma 3 and the cycle argument are completed rigorously.","tokens_in":19212,"tokens_out":10396,"duration_ms":97291,"concrete_test":"For a single obstacle with x*=(0,0), x1=(3,0), A1=diag(1,100), r1=1, compute the boundary curve ∂B_1^k by the equality in (34) for k=10, 100, 1000; parametrize the curve and use a root-finder to locate all zeros of the tangential component of g_new in (22) along it. At each zero, compute the derivative of the normal component along the curve to determine stability. If the number of zeros is greater than one or any zero is attracting, Lemma 3 is false; if only the aligned point appears and is unstable at every k, the conjecture is supported but the proof gap remains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, Theorem 1, requires that for k > K the trajectory exits every epsilon-neighborhood of each obstacle and visits only finitely many obstacles. Both requirements fail at the proof level. Lemma 3 (Appendix C) is introduced to guarantee a unique unstable critical point on the repulsion-zone boundary ∂B_i^k, and Lemmas 2-3 are what ensure escape from the repulsion zone. But the proof of Lemma 3 begins with 'We conjecture that the saddle point on the border of the repulsion zone will be close to...' and then only shows that a point satisfying (44) would be unstable. It never proves that such a point lies on ∂B_i^k, that it is a critical point of the restricted flow, or that no other critical points exist on ∂B_i^k; the sentence 'What is left to show is that there is no other critical point' is not followed by a derivation. Since g2 in (47) is only argued to be O(1/k) and 'dominated', the analysis of g1 does not cover the full vector field (22). Separately, Section IV-B's finite-visit argument for cyclic configurations asserts without proof that cycles require the obstacles to be equidistant from the objective and that infinite looping forces crossing the obstacle center; Lemma 2 only controls a normal velocity component at points away from alignment and does not imply those geometric claims. Thus the theorem's conclusion, and even the abstract's weaker almost-everywhere claim, are unsupported. The numerical experiments, while encouraging, do not establish the existential K for all configurations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modified gradient flow (Eq. 22) for navigating a point agent in a convex quadratic potential toward its minimizer while avoiding disjoint ellipsoidal obstacles. The modification applies a Hessian-inspired correction to the Rimon-Koditschek dynamics so that the agent only needs relative positions and distances, not curvature estimates. Theorem 1 claims that for a sufficiently large gain k the flow stays in the free space and converges to the target from every initial condition. The proof introduces global and local Lyapunov functions, a repulsion zone near each obstacle, and lemmas about escape from that zone and finite obstacle visits, followed by numerical comparisons with the uncorrected and second-order-corrected dynamics.","tokens_in":19493,"tokens_out":5662,"duration_ms":57089,"significance":"If Theorem 1 were established, the result would be practically significant: it would remove the eccentricity restriction (10) that limits Rimon-Koditschek navigation functions, and it would achieve this with a simple dynamics that does not require second-order information about the obstacles or the potential. The numerical experiments, especially the success-rate comparison as obstacle count increases, support the plausibility of the claim. However, the proof as written contains load-bearing gaps: a central lemma is justified by an explicit conjecture, and the finite-visit argument for cyclic configurations is asserted rather than proved. The paper is therefore not currently a rigorous proof of its main theorem.","major_comments":[{"comment":"The proof of Lemma 3 begins with the sentence 'We conjecture that the saddle point on the border of the repulsion zone will be close to the point xs' satisfying (44), and the subsequent analysis only shows that such a point, if it exists, is unstable. It is never shown that a point satisfying (44) lies on ∂B_i^k, that it is a critical point of the restricted flow, or that no other critical point exists on that boundary. The passage beginning 'What is left to show is that there is no other critical point' is followed by an appeal to Lemma 2 and a maximum over K_i^γ, not by a derivation that rules out other critical points. Since Lemma 3 is essential for the claim that the agent escapes the repulsion zone, Theorem 1 is not established by the arguments given.","section":"Appendix C (Lemma 3)"},{"comment":"The finite-visit argument for cyclic configurations is asserted without proof. The text states that 'For cycles to exist, it must be that the obstacles are equidistant from the objective' and that an infinite loop 'requires it to cross the center of the obstacle,' but neither statement is derived from the dynamics. Lemma 2 only establishes that a particular normal component of the velocity is positive when (x−xi) is not aligned with (x−x*); it does not imply the geometric claims about cycle structure or obstacle centers. Consequently, the conclusion that the agent visits only finitely many obstacles is unsupported, and this conclusion is needed for the convergence part of Theorem 1.","section":"Section IV-B"},{"comment":"Theorem 1 states convergence from every initial condition x0, while the abstract limits the claim to 'almost every initial position (up to a set of measure one)' and the proof text in Section IV-A says 'Lemmas 2 and 3 guarantee that the agent will exit the repulsion zone for almost every initial position.' These are materially different statements. The universal quantification in Theorem 1 is stronger than anything the proof establishes, so either the theorem must be weakened or the measure-zero caveat must be eliminated in the proof.","section":"Theorem 1 and abstract"},{"comment":"The proof of Lemma 1 derives, on the border of the repulsion zone, an expression that vanishes, but the lemma claims strict decrease in the interior and constancy on the boundary. The text says 'We have shown that the local Lyapunov function candidate is strictly negative except for on the border of the repulsion zone,' yet the displayed calculation only handles the boundary case. The interior strictness follows informally from the inequality in (33), but this step is not written out, and the notational shift between Vi in (31), V_i in (40), and the tilde function in (35) makes the argument harder to verify.","section":"Appendix A (Lemma 1)"}],"minor_comments":[{"comment":"There is a typo in 'the numerator on the right hand since of (26)': 'since' should be 'side.'","section":"Section IV, after Eq. (26)"},{"comment":"In the proof of Lemma 4, the text refers to 'the functions f0(x), B(x), and β(x) are continuous,' but B(x) is not defined in this paper and does not appear in the expression being bounded; the factor is sum_i \\barβ_i(x), so the reference to B(x) appears to be a remnant from the second-order dynamics.","section":"Appendix D, Eq. (56)"},{"comment":"Reference [31] has the same title and author list as the present manuscript and appears to be an unpublished companion or prior version; the paper should state its status and clarify how the present result differs from it, especially since the introduction compares the proposed dynamics with those of [31].","section":"Reference [31]"},{"comment":"The word 'centeres' should be 'centers' in 'The centeres are then chosen.'","section":"Section V-B"}],"recommendation":"reject","confidential_remarks":"The paper's central theorem is not proved: Lemma 3 is based on an explicit conjecture, and the finite-visit argument in Section IV-B is a geometric assertion without derivation. These are not routine details but load-bearing components of the convergence proof. The numerical results are encouraging, and the idea of a Hessian-inspired correction is interesting, but the current manuscript does not meet the standard for publication as a rigorous proof-based paper. If the authors can supply a complete proof of Lemma 3 and a rigorous finite-visit argument, a resubmission would be worth considering. The relationship to reference [31] should also be clarified, as that reference appears to be the same title and authors, and the claimed improvement over it is part of the paper's stated contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I agree with the stress-test reading, and I think the reader's verdict is right. The short version: this paper has a genuinely attractive idea, namely replacing the second-order RK correction with a simple sum of relative-position terms, and the simulations show it works in worlds where condition (10) fails. But the proof of Theorem 1 has two load-bearing gaps, and the theorem as stated is not established.\n\nWhat is new: the dynamics (22) themselves, which require only distances and relative positions, and the claimed guarantee for all initial positions in ellipsoidal environments with arbitrarily flat obstacles. That is stronger than the companion paper [31], which needs second-order information and an epsilon-away start. The numerical study is extensive for a theory paper: comparisons against standard RK and second-order corrections, 3-D examples, and random worlds with up to seven obstacles. I believe the empirical evidence: the corrected field does avoid the local minima that plague (12).\n\nWhere it breaks. Lemma 3's proof in Appendix C begins 'We conjecture that the saddle point on the border of the repulsion zone will be close to...'. Then it analyzes that conjectured point and shows it is unstable, but it never proves the point is on ∂B_i^k, that it is a critical point, or that no other critical point exists there. The sentence 'What is left to show is that there is no other critical point' is followed by a gamma-neighborhood argument that assumes the very uniqueness it needs. Since escape from the repulsion zone hinges on Lemma 3, this is not a small gap. Section IV-B's cycle argument is also asserted: 'For cycles to exist, it must be that the obstacles are equidistant from the objective' and that infinite looping forces crossing the obstacle center. Lemma 2 only gives a normal-velocity sign condition in a hyperplane; it does not imply those geometric claims. So the all-initial-conditions guarantee, and even the abstract's 'almost every' in the header, are unsupported. The paper is honest about the conjecture, which I appreciate, but an honest conjecture inside a proof of the main theorem is still a missing proof.\n\nMinor: the saturation constant K is existential, never characterized; for an implementation-oriented paper, that leaves a tuning parameter entirely unspecified. And the abstract is inconsistent, 'every initial position' versus 'almost every', which probably reflects the proof gap.\n\nBottom line: the idea deserves attention and the numerics are real, but this needs a serious rewrite of the proof. A weaker theorem with a complete proof, or a genuine proof of Lemma 3 and the cycle claim, would change my verdict. I would send it to peer review, with a clear request for major revision.","headline":"A simple and appealing correction to RK navigation dynamics, but the main theorem is unproven: the proof leans on a conjectured critical point and an unproved cycle argument.","tokens_in":20037,"tokens_out":3964,"would_cite":false,"duration_ms":36820,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a Hessian-corrected gradient flow navigates a convex quadratic potential around arbitrarily flat ellipsoidal obstacles from any free-space start.","keywords":["quadratic potential","ellipsoidal obstacles","navigation functions","gradient flow","Hessian correction","Lyapunov stability","artificial potential fields","motion planning"],"falsifier":"Simulate (22) with large $k$ on a symmetric two-obstacle world with the target between the obstacles, and numerically solve $\\dot x=0$ over each repulsion zone $\\partial B_i^k$; finding a second critical point or a stable node inside any repulsion zone would disprove Lemma 3. Separately, place several obstacles exactly equidistant from the target in a ring and test whether any trajectory under exact arithmetic revisits the same obstacle repeatedly, which would disprove the finite-visit claim.","tokens_in":18963,"feed_emoji":"🧭","tokens_out":8940,"duration_ms":85342,"temperature":0.7,"pith_summary":"This paper claims that a point agent can always reach the minimum of a convex quadratic potential in a workspace punctured by disjoint ellipsoidal obstacles by following a modified gradient flow, no matter how flat the obstacles are. The modification multiplies each gradient in a Rimon-Koditschek navigation flow by the inverse Hessian of the function defining it, which for quadratics simplifies to moving toward the goal and away from each obstacle center, so no curvature measurements are needed in practice. The main theorem states that for a sufficiently large gain $k$, trajectories starting in free space stay in free space and converge to the target. The paper also shows numerically that the correction succeeds in worlds where the standard Rimon-Koditschek potential provably fails. What is at stake is replacing a geometric restriction on obstacle shape with a finite, start-independent tuning knob.","feed_headline":"Flat ellipsoids stop blocking gradient navigation","feed_subtitle":"A curvature-corrected navigation flow steers around flat ellipsoids without curvature measurements.","key_machinery":"The load-bearing object is the curvature-corrected navigation flow (22). It is a gradient-like dynamical system in which each repulsive obstacle term is weighted by the omitted product $\\bar\\beta_i(x)$ and points from the ellipsoid center $x_i$, while the attractive term points from the goal $x^*$. The argument around it uses a global Lyapunov function $V(x)=\\frac12\\|x-x^*\\|^2$, a per-obstacle Lyapunov function $V_i(x)=\\frac12(x-x^*)^\\top\\nabla^2\\beta_i(x)(x-x^*)$, and the repulsion zone $B_i^k(\\delta)$, the thin region where the candidate Lyapunov functions fail to decrease. The key mechanism is that increasing $k$ shrinks the repulsion zones into $\\varepsilon$-balls around obstacles, within which the flow points away from obstacle centers, forcing escape.","core_discovery":"The central claim is Theorem 1: for $f_0(x)=(x-x^*)^\\top Q(x-x^*)$ and ellipsoidal obstacles $\\beta_i(x)=\\frac12(x-x_i)^\\top A_i(x-x_i)-\\frac12 r_i^2$, there is a finite threshold $K$ such that for all $k>K$ the flow $\\dot x=-\\beta(x)(x-x^*)+\\frac{f_0(x)}{k}\\sum_i \\bar\\beta_i(x)(x-x_i)$ keeps every free-space trajectory inside the free space and sends it to $x^*$. The flow is obtained by replacing $\\nabla f_0$ and $\\nabla\\beta_i$ with Hessian-inverse products, which reduce to relative-position vectors; this recreates the dynamics of a spherical world without assuming the world is spherical. The proof splits the state space: far from obstacles a global Lyapunov function decreases; near each obstacle local Lyapunov functions define a repulsion zone that the agent leaves; and a graph-ordering argument limits the number of obstacles visited. The theorem is stated for every initial free-space position, though the proof's local lemmas establish escape only up to a measure-zero set, a gap the paper does not reconcile.","pith_inferences":["The same flow should extend to disjoint strictly convex obstacles by locally replacing each obstacle with an inscribed ellipsoid or a quadratic barrier, at the cost of reintroducing shape estimation.","The no-second-order-information feature is special to quadratic potentials and ellipsoidal barriers; for a general strongly convex potential, $\\nabla^2 f_0^{-1}\\nabla f_0$ is not $x-x^*$, so the simplification would not survive without curvature measurements.","The proof's measure-zero caveat predicts that non-converging initial conditions, if they exist, lie on stable manifolds of repulsion-zone saddles; a numerical search initialized exactly on those manifolds could test whether the 'every initial position' form of Theorem 1 is actually necessary."],"forward_implications":["If Theorem 1 is correct, the eccentricity restriction on Rimon-Koditschek navigation disappears for quadratic costs and ellipsoidal obstacles: condition (10) no longer needs to hold.","Implementation requires only relative positions and distances to the goal and obstacle centers, because the Hessian corrections collapse into $x-x^*$ and $x-x_i$.","The threshold $K$ is finite and independent of the starting point, so one gain setting works for all initial conditions in a given ellipsoidal world.","The simulations show that success rates stay above 95 percent for $k=60$ as the obstacle count grows, while uncorrected navigation degrades; the theorem explains this by removing the old curvature condition."],"supporting_citations":[{"why":"Supplies the definition of navigation functions and the general guarantee that following their negative gradient reaches the target for almost every start.","marker":"[14]"},{"why":"Establishes that Rimon-Koditschek potentials are navigation functions in sphere worlds, the baseline that the proposed correction imitates.","marker":"[15]"},{"why":"Gives the sufficient-curvature restriction that limits when Rimon-Koditschek potentials work, which motivates the correction.","marker":"[20]"},{"why":"Supplies condition (10), the eccentricity restriction Theorem 1 removes, and the convex-potential navigation-function framework the paper builds on.","marker":"[21]"},{"why":"Provides the convex-analysis tools, including Hessian-inverse gradients and separating hyperplanes, used to construct the corrected flow and graph argument.","marker":"[30]"},{"why":"Presents the earlier second-order-corrected dynamics that this paper simplifies; its initial-position-dependent threshold and cycle limitation are the comparison points.","marker":"[31]"}],"fun_headline_variants":["Hessian fix unblocks flat-obstacle navigation","Curvature correction steers past flat ellipsoids","Ellipsoid eccentricity no longer blocks navigation","Gradient trick navigates any ellipsoid obstacle","Modified flow navigates flat ellipsoids robustly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's load-bearing premise is that each repulsion zone has exactly one critical point, an unstable saddle near the line through $x^*$ and $x_i$, and that cycles are escaped because infinite looping would force crossing an obstacle center; both are asserted rather than proven, one explicitly as a conjecture, and the convergence claim collapses if either fails.","fun_headline_variants_meta":{"raw":{"variants":["Hessian fix unblocks flat-obstacle navigation","Curvature correction steers past flat ellipsoids","Ellipsoid eccentricity no longer blocks navigation","Gradient trick navigates any ellipsoid obstacle","Modified flow navigates flat ellipsoids robustly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1557,"prompt_tokens":986,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":494}},"tokens_in":602,"tokens_out":571,"duration_ms":5596,"temperature":1.0,"reasoning_tokens":494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:37:48.760341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate (22) with large $k$ on a symmetric two-obstacle world with the target between the obstacles, and numerically solve $\\dot x=0$ over each repulsion zone $\\partial B_i^k$; finding a second critical point or a stable node inside any repulsion zone would disprove Lemma 3. Separately, place several obstacles exactly equidistant from the target in a ring and test whether any trajectory under exact arithmetic revisits the same obstacle repeatedly, which would disprove the finite-visit claim.","supporting_citations":[{"cited_title":"Robot navigation functions on manifolds with boundary,","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of navigation functions and the general guarantee that following their negative gradient reaches the target for almost every start."},{"cited_title":"Exact robot navigation using artificial potential functions,","cited_arxiv_id":null,"evidence_quote":"Establishes that Rimon-Koditschek potentials are navigation functions in sphere worlds, the baseline that the proposed correction imitates."},{"cited_title":"Navigation functions for every- where partially sufficiently curved worlds,","cited_arxiv_id":null,"evidence_quote":"Gives the sufficient-curvature restriction that limits when Rimon-Koditschek potentials work, which motivates the correction."},{"cited_title":"Navigation functions for convex potentials in a space with convex obstacles,","cited_arxiv_id":null,"evidence_quote":"Supplies condition (10), the eccentricity restriction Theorem 1 removes, and the convex-potential navigation-function framework the paper builds on."},{"cited_title":"Navigation of a quadratic potential with ellipsoidal obstacles","cited_arxiv_id":null,"evidence_quote":"Presents the earlier second-order-corrected dynamics that this paper simplifies; its initial-position-dependent threshold and cycle limitation are the comparison points."}],"review_version":1}