{"id":"92db8fcf-dded-4465-b230-9eb38ac498ca","arxiv_id":"2608.02834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"BMTP is a biconvex minimum-time planner that convexifies time and derivative constraints jointly, replaces convex decomposition with time-varying separating planes, and guarantees monotone duration decrease on every feasible iterate.","lead":"This paper introduces a motion-planning algorithm that computes fast, collision-free robot trajectories by alternating between optimizing the path and optimizing walls that keep it away from obstacles. It needs no precomputed map of the free space, respects smoothness limits such as jerk and snap, and matches or beats existing planners in drone and warehouse tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence guarantee does not transfer to the implemented finite-dimensional planner: the plane update is not shown to produce the Delta_v-strict separation needed for monotone duration, and the fixed plane degree can make the Sec. VI-D assumption fail.","rationale":"Rationale: The paper has real strengths. Lemma 1, the polar characterization, and the maximum-margin plane derivation in App. D are mathematically sound as far as I verified; the change of variables T_I = T^I is valid and correctly convexifies the minimum-time objective and derivative constraints; the open-source release and hardware validation are meaningful. The weakest point is the transfer from the idealized plane update in Sec. VI-D to the finite-dimensional algorithm in Sec. VII. The reader's weakest_assumption identified the same issue: App. D only establishes nonpositive optimal separation for the surrogate, while the trajectory update requires the strict margin -Delta_v. I agree this is the most load-bearing concern because the abstract's unconditional 'guaranteed to converge' and the anytime property both depend on monotone non-increase of the duration across outer iterations, which in turn depends on the previous feasible r remaining feasible after every plane update. The fixed plane degree is a second, independent route to the same failure: the theorem assumes plane-update feasibility for every non-colliding segment-obstacle pair, and later iterates are Bezier curves rather than polygonal segments, so the statement that the assumption holds for the initial polygonal curve does not carry over. These issues do not invalidate the empirical comparisons or the practical value of the algorithm, but they mean the formal guarantee is conditional in a way not conveyed by the abstract. The proposed synthetic test is small and would settle whether the monotone property actually holds in the released implementation. If it fails, the paper must either prove a margin-aware plane update or weaken the claimed guarantee; the appropriate verdict remains CONDITIONAL rather than REJECT because the method may still be correct and useful once the claim is scoped properly.","tokens_in":26204,"tokens_out":16370,"duration_ms":156257,"concrete_test":"Run BMTP on a 2D instance with one convex obstacle, using the default Delta_p = Delta_v = 1e-6 and plane degree 1. Initialize with a collision-free degree-8 Bezier segment whose closest approach to the obstacle is d = 5e-7 (below Delta_v but collision-free). After the first feasible trajectory update, record the duration T; then execute one outer iteration exactly as in Alg. 1 (recompute planes around the new feasible r, then call TRAJECTORYUPDATE). Check whether the previous r satisfies the new constraint v_{m,k}(s) <= -Delta_v. If it does not, record whether TRAJECTORYUPDATE returns infeasible or a feasible candidate with duration greater than T. Either outcome falsifies the monotone-duration lemma on which the convergence and anytime guarantees rest. Rerun with d = 1e-5 as a control to confirm the planner behaves normally when the margin assumption holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Load-bearing concern: the central formal claim is that BMTP is guaranteed to converge and is anytime (abstract, Sec. I-C). Theorem VI-D proves this for an idealized plane update that is feasible for every non-colliding segment-obstacle pair and whose returned planes keep the previous feasible trajectory in the feasible set of the next trajectory update. The implemented plane update (Sec. VII-C) and trajectory update (Sec. VII-D) are not shown to satisfy this condition. Appendix D proves only that the surrogate's optimal separation value is nonpositive for a collision-free r; the trajectory update then requires v_{m,k}(s) <= -Delta_v with Delta_v = 1e-6. If the current feasible r comes within less than Delta_v of an obstacle, no unit-norm separating plane can satisfy that strict bound, so the previous iterate leaves the feasible set and the monotone-duration argument collapses. Even without the margin, the plane functions are restricted to fixed-degree Bezier curves (degree 1 in all experiments), whereas the theorem permits unrestricted plane functions; the paper's own limitation section concedes there is no principled way to choose the plane degree. A curved collision-free segment can fail to be separable by a degree-1 plane, so the feasibility assumption is not established for later Bezier iterates. Thus the unconditional 'guaranteed to converge' overstates what is proved for the actual algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents BMTP, a biconvex optimization approach for minimum-time trajectory planning around convex obstacles with derivative constraints of arbitrary order. The authors jointly convexify the minimum-time objective and derivative constraints via a change of variables (credited to [42]), reformulate collision avoidance through time-varying separating planes drawn from the polar of each obstacle, and alternate between a maximum-margin plane update and a convex trajectory update. A finite-dimensional implementation uses composite Bézier curves for the trajectory and the separating planes, with conservative control-point enforcement of separation and collisions. The paper claims a convergence guarantee and an anytime property under a feasibility assumption, and demonstrates the method in simulation and on hardware for a quadrotor village navigation task and a dual-arm bin-unloading task, comparing against a decomposition-based planner (EI+SCS) and a nonlinear-programming baseline. An open-source implementation is provided.","tokens_in":26364,"tokens_out":11536,"duration_ms":104288,"significance":"If the claims held, this would be a substantial advance: a decomposition-free, anytime planner with arbitrary-order derivative constraints, strong empirical robustness to initialization, open-source code, and hardware validation. The paper's derivations of the convexification (Lemma 1), the polar recipe (Appendix B), and the maximum-margin plane dual (Appendix D) are careful and largely correct, and the empirical evaluation is extensive (300 randomized instances plus hardware trials). The central theoretical guarantee, however, is not established for the implemented finite-dimensional algorithm, which is a load-bearing issue that affects the main claims in the abstract and Section I-C.","major_comments":[{"comment":"The convergence proof in Sec. VI-D relies on the previous feasible trajectory remaining in the feasible set of the next trajectory update. The implemented trajectory update (Sec. VII-D, Eq. (13)) imposes v_{m,k}(s) <= -Delta_v with Delta_v = 1e-6, while App. D establishes only that, for a collision-free r, the surrogate (12) has optimal separation value 'nonpositive' (v <= 0), not v <= -Delta_v. If r passes within distance less than Delta_v of an obstacle, no unit-norm separating plane can satisfy (13), so the previous iterate leaves the feasible set and the monotone-duration argument collapses. The margin Delta_v is not a minor technicality: it is the entire slack between the proven statement and the required one.","section":"Sec. VI-D, Sec. VII-D, App. D"},{"comment":"The theorem in Sec. VI-D assumes the plane update is feasible for every non-colliding segment-obstacle pair, but the implemented plane update restricts the plane functions to fixed-degree Bézier curves (degree 1 in all experiments, Secs. VII-C and VIII). A collision-free curved segment need not be separable by a degree-1 plane function; for example, an arc around a disk requires the separating normal to rotate nonlinearly with s, so a restricted plane family can fail. App. D's assertion that the discrete program retains the feasibility property at v <= 0 is not proven and is false in general. The paper itself concedes in Sec. IX(d) that there is no principled way to select the plane degree. Hence the Sec. VI-D assumption is not established for the actual algorithm, and the unconditional 'guaranteed to converge' in the abstract overstates what is proved.","section":"Sec. VI-D, Sec. VII-C, Sec. IX(d)"},{"comment":"The inner-loop termination bound ('at most KM attempts') assumes that once a pair (k,m) is tagged, the pair can never collide again. This property is a consequence of the same strict-separation guarantee that is missing in the finite-dimensional implementation; if a tagged pair re-collides, the bound no longer applies, and the anytime claim, which depends on monotone duration decrease, also fails. The proof should either be extended to the finite-dimensional setting or the statements in the abstract and Sec. I-C should be qualified.","section":"Sec. VI-D"}],"minor_comments":[{"comment":"The notation T_I and the fractional exponent T_I^{i/I} is difficult to parse in the typeset text; please define T_I explicitly as the I-th power of T and use a clearer format.","section":"Sec. IV, Eq. (5)"},{"comment":"The sentence 'we split the plane update O_k into independent subproblems' should read 'we split the plane-update problem for obstacle O_k into independent subproblems.'","section":"Sec. VII-C"},{"comment":"The reported FPP computation times should state explicitly whether the 1.64 s preprocessing cost is included; Table I currently lists times that appear to exclude it.","section":"Sec. VIII-A, Table I"},{"comment":"The limitation list does not mention the gap between the idealized convergence proof and the finite-dimensional plane update; adding a sentence acknowledging that the guarantee holds only under the Sec. VI-D feasibility assumption would improve accuracy.","section":"Sec. IX(d)"},{"comment":"The sentence 'the constraint (16d) is redundant since there are no continuity constraints on (a,b)' is misleading for the finite-dimensional problem, where the restricted plane family couples the variables across s; please justify the redundancy by the pointwise optimality argument or remove it.","section":"App. D, after Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The paper is from a strong group and the empirical work is impressive. My main concern is the mismatch between the claimed convergence guarantee and what is actually proved for the implemented algorithm; this should be resolved before publication. The authors have the tools to fix it (e.g., by adding an adaptive plane-degree or segment-refinement step, or by clearly stating the additional assumption under which the guarantee holds), so I recommend major revision rather than rejection. I would also ask the editor to ensure the claims in the abstract are matched to the algorithm actually implemented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the new content is the time-varying separating-plane formulation with obstacle tagging, together with the joint convexification of the minimum-time objective and arbitrary-order derivative constraints. I checked Lemma 1, the polar recipe in App. B, and the max-margin derivation in App. D; they're correct. The biconvex alternation is described clearly, and the experiments—village drone, bin unloading, hardware—look honest and reproducible in spirit. The open-source implementation helps.\n\nThe main soft spot is the convergence proof. Theorem VI-D proves monotone duration and finite inner loop for an idealized plane update that keeps the previous feasible trajectory strictly feasible. The implemented plane update (Sec. VII-C) uses fixed-degree Bézier planes and enforces a margin Δ_v=1e-6. Appendix D only shows that the surrogate's optimal separation value is nonpositive for a collision-free r. If r comes within Δ_v of an obstacle, the max-margin plane does not satisfy the strict bound v ≤ -Δ_v, so the previous iterate leaves the feasible set and the monotone-duration argument collapses. And nothing guarantees that a degree-1 plane can separate a curved segment from a convex obstacle; the paper's own limitations section says there is no principled way to choose the plane degree. So the abstract's 'guaranteed to converge' overstates what is proved for the actual algorithm. This is fixable by restating the theorem with the margin and plane degree as explicit assumptions, and by reporting what happens when the plane subproblem is infeasible.\n\nSmaller issues: the village 'lower bound' of 11.35 s is a valid relaxation, but the 'only 0.48 s' framing flatters the obstacle cost; fine as a bound but should be labeled as the gap to the obstacle-free problem. The robustness-to-initialization comparison in Tab. III leaves out EI+SCS under the sequential initialization, so the 'more robust' claim lacks its main comparison. No commit hash or seeds, so independent reproduction takes effort.\n\nNone of this makes the method a bad paper. The biconvex formulation and tagging mechanism are genuinely new and useful. But the formal claim needs to match the numerics. I'd send it to review with a request for major revision on the convergence statement, the missing baseline, and versioned code. A serious referee should definitely see it.","headline":"Solid biconvex motion planner with a real gap between the proved convergence claim and the implemented algorithm; worth reviewing, needs a revised guarantee.","tokens_in":27049,"tokens_out":5624,"would_cite":true,"duration_ms":48286,"reading_group":"yes","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 proposes BMTP, a biconvex minimum-time planner that guarantees convergence and is anytime while supporting arbitrary-order derivative and continuity constraints, with no upfront convex decomposition of the free space.","keywords":["biconvex optimization","minimum-time trajectory","motion planning","convex obstacles","separating planes","derivative constraints","Bezier curves","anytime planning"],"falsifier":"Run BMTP with the margins $\\Delta_p$ and $\\Delta_v$ set to zero on an instance where a trajectory segment runs tangent to a convex obstacle, so the maximum-margin separating plane has zero separation at the tangency point; if the inner loop then fails to terminate within $KM$ tags or the duration ever increases, the plane-update feasibility assumption is false.","tokens_in":25794,"feed_emoji":"🤖","tokens_out":10111,"duration_ms":82285,"temperature":0.7,"pith_summary":"This paper introduces BMTP, a biconvex minimum-time motion planner for robots moving among convex obstacles — biconvex meaning that the problem is convex in the trajectory when the separating planes are fixed, and convex in the planes when the trajectory is fixed. It claims to guarantee convergence from any collision-free polygonal curve, to be anytime (an interrupted run returns the best feasible trajectory found so far), and to enforce derivative constraints of arbitrary order — velocity, acceleration, jerk, snap — plus continuity to arbitrary finite degree, without a costly upfront convex decomposition of the free space. The key move is a change of variables that convexifies the minimum-time objective and all derivative constraints, leaving collision avoidance as the only nonconvexity; that residual nonconvexity is handled by time-varying separating planes optimized alternately with the trajectory, adding planes only for obstacles the current iterate collides with. In experiments on quadrotor navigation through a 521-obstacle village and on dual-arm bin unloading in simulation and on hardware, BMTP matches the trajectory quality and computation time of a state-of-the-art decomposition-based planner while handling a larger class of problems and being less sensitive to poor initialization.","feed_headline":"Biconvex planner converges to min-time paths around convex obstacles","feed_subtitle":"Supports jerk and snap limits, needs no convex decomposition, and returns anytime feasible trajectories.","key_machinery":"The load-bearing object is the time-varying separating plane drawn from the polar $O_k^\\circ = \\{(a,b) : a^\\top x + b \\ge 0 \\ \\forall x \\in O_k\\}$ of each convex obstacle, together with the change of variables $T_I = T^I$. The polar is always convex, so fixing either the trajectory or the planes leaves a convex subproblem; the bilinear terms $a_k(s)^\\top r(s)$ are the only remaining nonconvexity. The change of variables turns derivative constraints $r^{(i)}(s) \\in T^i C_i$ into $r^{(i)}(s) \\in T_I^{i/I} C_i$, which are convex because $T_I^{i/I}$ is concave in $T_I$ and each $C_i$ contains the origin. The plane update computes a maximum-margin separating plane by projecting the current trajectory point onto the obstacle, with a closed form in the ideal infinite-dimensional case and a conservative Bezier-surrogate version in the finite-dimensional implementation. The convex-hull property of Bezier curves lets derivative and plane constraints be enforced at control points, keeping the trajectory update convex, and tagging only the obstacle—segment pairs that currently collide gives the procedure a mechanism to jump around obstacles and escape local minima.","core_discovery":"The paper's central claim is that the minimum-time trajectory problem of (1) can be reformulated as a biconvex program that is guaranteed to converge and is anytime. The reformulation normalizes time to $s \\in [0,1]$, works with the variable $T_I = T^I$ so that derivative constraints become $r^{(i)}(s) \\in T_I^{i/I} C_i$, which are convex by Lemma 1, and the objective becomes monotone in $T_I$. Collision avoidance is rewritten as a search for time-varying separating planes $(a_k(s), b_k(s)) \\in O_k^\\circ$ satisfying $a_k(s)^\\top r(s) + b_k(s) < 0$ for all $s$, where $O_k^\\circ$ is the polar of the obstacle. The resulting problem is convex in the trajectory given the planes, and convex in the planes given the trajectory. BMTP alternates between a plane update (projecting the current feasible trajectory onto each tagged obstacle to get maximum-margin supporting planes) and a trajectory update (convex optimization with planes fixed), adding planes only for obstacles the current iterate collides with; the paper proves the inner loop terminates after at most $KM$ tags and the duration never increases across outer iterations, so every feasible iterate is a valid anytime solution. Empirically, on a 521-obstacle village drone problem it computes a snap-constrained 11.83 s trajectory in 0.19 s, and on 300 randomized dual-arm bin unloading instances it matches the decomposition-based EI+SCS baseline in duration (2.82 s vs 2.77 s average) and computation time (188 ms vs 204 ms), with 100% success and collision-free plans on both, while also tolerating weaker initializations.","pith_inferences":["An implicit consequence the paper does not draw out: the shared segment-duration formulation is what makes continuity constraints linear; allowing per-segment durations, which decomposition-based planners typically use, would reintroduce nonlinear equality constraints, so the uniform-timing assumption is a structural limit rather than a practical convenience.","The independence of plane-update subproblems across segment—obstacle pairs suggests a straightforward GPU parallelization, which could push the reported 188 ms computation time down further and make the method attractive for high-rate replanning; the paper only parallelizes the collision checker.","The same convexification could be applied to polish teleoperated demonstration trajectories: feeding a waypoint path through BMTP would tighten wide berths while preserving a user's rough route, potentially improving imitation-learning datasets without changing the policy class.","A testable extension would replace the conservative control-point bound on $v_{m,k}$ with sum-of-squares certification, which the paper notes is exact but currently too expensive; if the cost drops, the strict-margin assumption could be certified rather than left to the $10^{-6}$ margins."],"forward_implications":["BMTP can be used as an anytime planner: interruption at any outer iteration returns a collision-free trajectory whose duration never exceeds that of any earlier feasible iterate.","Because derivative and continuity constraints of arbitrary order are convexified, minimum-time trajectories can be planned directly with jerk and snap limits, without heuristic smoothing or retiming post-processing.","The method eliminates the need to precompute a convex decomposition of the free space; when the environment changes, only the separating planes need to be recomputed on the fly around the current trajectory.","The tagging mechanism lets the trajectory escape poor initialization, cutting diagonally through a cluttered village despite a naive perimeter route, which a planner restricted to a fixed safe corridor cannot do.","On bin unloading, BMTP matches a strong decomposition-based baseline in solution quality and runtime while maintaining 100% success and collision-freedom across 300 randomized instances."],"supporting_citations":[{"why":"Supplies the change of variables $T_I = T^I$ used to convexify the minimum-time objective and all derivative constraints.","marker":"[42]"},{"why":"Provides the convex-analysis background (Lemma 1 on sets scaled by concave functions) and the conic polar duality used in the plane formulation.","marker":"[43]"},{"why":"The decomposition-based minimum-time planner through convex sets that serves as the main baseline (EI+SCS) and supplies the per-segment retiming used for waypoint initialization.","marker":"[3]"},{"why":"Provides the village environment and the Fast Path Planning baseline that BMTP is compared against in the drone experiment.","marker":"[28]"},{"why":"Supplies the edge-inflation algorithm used to generate the safe-set sequence for the EI+SCS baseline.","marker":"[29]"},{"why":"The interior-point solver used to solve the convex trajectory-update subproblems and the per-segment retiming programs.","marker":"[53]"},{"why":"The survey of biconvex optimization that motivates the alternating plane-update/trajectory-update scheme.","marker":"[44]"},{"why":"The convex-concave-like tagging mechanism the planner uses to escape local minima by only adding planes for colliding obstacles.","marker":"[46]"}],"fun_headline_variants":["Biconvex alternation guarantees min-time paths, supports jerk/snap","Anytime min-time planner with no convex decomposition","Guaranteed-convergence min-time planner, robust to bad init","Alternating planes give guaranteed min-time trajectories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Convergence assumes that for every segment—obstacle pair that does not collide, the plane update returns planes that keep the current feasible trajectory strictly separated from the obstacle by a tiny prescribed margin ($10^{-6}$), whereas the appendix only proves the surrogate's optimal separation is nonpositive for a collision-free path.","fun_headline_variants_meta":{"raw":{"variants":["Biconvex alternation guarantees min-time paths, supports jerk/snap","Anytime min-time planner with no convex decomposition","Guaranteed-convergence min-time planner, robust to bad init","Alternating planes give guaranteed min-time trajectories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000888,"raw_usage":{"total_tokens":3902,"prompt_tokens":1086,"completion_tokens":2816,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":702,"completion_tokens_details":{"reasoning_tokens":2747}},"tokens_in":702,"tokens_out":2816,"duration_ms":19676,"temperature":1.0,"reasoning_tokens":2747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:02:53.269150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BMTP with the margins $\\Delta_p$ and $\\Delta_v$ set to zero on an instance where a trajectory segment runs tangent to a convex obstacle, so the maximum-margin separating plane has zero separation at the tangency point; if the inner loop then fails to terminate within $KM$ tags or the duration ever increases, the plane-update feasibility assumption is false.","supporting_citations":[{"cited_title":"Time-optimal control of a multidimensional integrator chain with applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the change of variables $T_I = T^I$ used to convexify the minimum-time objective and all derivative constraints."},{"cited_title":"A biconvex method for minimum-time motion planning through sequences of convex sets,","cited_arxiv_id":null,"evidence_quote":"The decomposition-based minimum-time planner through convex sets that serves as the main baseline (EI+SCS) and supplies the per-segment retiming used for waypoint initialization."},{"cited_title":"Clarabel: An interior-point solver for conic programs with quadratic objectives,","cited_arxiv_id":null,"evidence_quote":"The interior-point solver used to solve the convex trajectory-update subproblems and the per-segment retiming programs."},{"cited_title":"Biconvex sets and optimiza- tion with biconvex functions: a survey and extensions,","cited_arxiv_id":null,"evidence_quote":"The survey of biconvex optimization that motivates the alternating plane-update/trajectory-update scheme."},{"cited_title":"Variations and extension of the convex–concave procedure,","cited_arxiv_id":null,"evidence_quote":"The convex-concave-like tagging mechanism the planner uses to escape local minima by only adding planes for colliding obstacles."}],"review_version":1}