{"id":"cee30fc9-5c36-4d7f-9a4d-45ad0206ae1f","arxiv_id":"2509.07239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A perception-informed dynamic gap planner that propagates predicted gaps forward in time and uses pursuit guidance to generate provably collision-free local trajectories under ideal conditions.","lead":"This paper introduces a gap-based local planner that tracks regions of free space in real time and predicts how they move, so a robot can pick a safe path through moving obstacles. It adds a formal collision-free guarantee under ideal assumptions and shows the planner outperforming eight baseline planners in simulated dynamic environments and on a real TurtleBot2.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof is unsupported: Eq. 22 applies the gap-goal speed ratio K to the left/right gap points whose speeds generally differ, and an interval bound on θ_e does not imply the robot's actual intercept heading γ_e lies between the endpoint intercept headings.","rationale":"The reader correctly identifies the constant-velocity assumption as a practical limitation, and the paper's own §5.5.3 admits that few gaps satisfy it. However, the more acute issue for the paper's central formal claim is that Theorem 1's proof has an internal logical gap independent of how often the assumptions hold. The proof's interval-comparison argument conflates the speed ratio for the gap goal with the speed ratios for the individual gap points and then treats a bound on θ_e as a bound on the robot's inertial intercept heading. Even if the theorem's conclusion is true (and it may be, since the convexity of p_g(t) gives a simpler route to the conclusion), the written proof does not establish it. This is a correctness risk in the core contribution, not merely a scope limitation. The experimental validation in §6.1 provides empirical support but does not repair the proof. The verdict should remain CONDITIONAL: acceptance should require either a corrected proof of Theorem 1 or an explicit restriction of the claimed guarantee to cases where the proof's inference is valid, e.g., equal endpoint speeds and compatible β/γ intervals.","tokens_in":26339,"tokens_out":19265,"duration_ms":266945,"concrete_test":"Run a parameter search or SMT check over constant-velocity two-point gaps satisfying the theorem's hypotheses and Eq. 15–16 for p_g, with ||v_l|| ≠ ||v_r||. For each sample, compute θ_e from Eq. 22 with K = v_e/v_g, and compute the true endpoint intercept headings θ*_{e/r}, θ*_{e/l} using K_l = v_e/||v_l|| and K_r = v_e/||v_r||. Check whether θ_e ∈ [θ*_{e/r}, θ*_{e/l}] and γ_e = θ_e + β_g ∈ [γ*_{e/r}, γ*_{e/l}]. Then simulate the corresponding PN trajectory against moving disks. If a sample violates the interval inclusion, the proof is invalid as written; if such a sample also produces a collision, Theorem 1 itself is false.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The formal guarantee in Theorem 1 (§5.5.3) rests on a proof step that does not follow. Eq. 15 defines K := v_e / v_g for the gap goal, and the proof uses monotonicity of arcsin to conclude that θ_e lies between θ_{e/r} and θ_{e/l}, described as 'the bearings at which the ego-robot must direct its velocity to intercept the left and right gap points.' But the endpoint targets have speeds ||v_l|| and ||v_r||, which generally differ from v_g; their interception headings must be computed with K_l = v_e/||v_l|| and K_r = v_e/||v_r||, not with K. Even if an interval bound on θ_e were valid, the robot's actual velocity direction is γ_e = θ_e + β_g, whereas the endpoint intercept directions are γ_{e/r} = θ_{e/r} + β_r and γ_{e/l} = θ_{e/l} + β_l. Summing the θ interval to a different β interval does not place γ_e between the endpoint directions. The conclusion 'therefore performing collision-free gap passage' is not established by the written argument. The theorem may be repairable — convexity of the convex combination p_g(t) might suffice — but the proof as published has a genuine logical gap in its central inference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents \"dynamic gap,\" a perception-informed gap-based local planner for dynamic environments. The planner detects and tracks polar free-space gaps, estimates their dynamics with a Kalman filter in the egocentric frame, propagates gaps forward to form gap tubes, uses parallel-navigation (PN) guidance from pursuit theory to determine gap feasibility and generate trajectories, and adds \"ungap\" planning when no gap exists. The central theoretical claim is Theorem 1: under ideal conditions—first-order holonomic ego-robot, constant-velocity gap points, and an isolated local environment—the PN policy toward a convex-combination gap goal yields collision-free gap passage. The paper also reports extensive simulation benchmarking in Arena-Rosnav against classical, learned, and prior gap-based planners, plus hardware experiments on a TurtleBot2. The authors state clearly that in practice few gaps satisfy the ideal assumptions, so the formal guarantee is conditional and supplementary modules provide robustness.","tokens_in":26748,"tokens_out":6777,"duration_ms":83080,"significance":"If Theorem 1 is valid, the paper supplies one of the few formal collision-free guarantees for gap-based planning in dynamic settings, complementing the typically emergent robustness of learned planners. The practical contributions are substantial: an open-source implementation, a full perception-space pipeline including gap tracking, propagation, gap tubes, ungap planning, and dynamic scan propagation, and a broad benchmarking study in three environments with classical and learned baselines. The Monte Carlo experiment in Section 6.1 is a useful self-consistency check, and the hardware deployment strengthens the empirical case. However, the formal contribution is currently undermined by a genuine gap in the proof of Theorem 1, so the theoretical significance is not yet established.","major_comments":[{"comment":"Theorem 1's proof is not established as written. The argument that θ_e ∈ [θ_{e/r}, θ_{e/l}] uses a single speed ratio K = v_e/v_g for all three interception headings. The left and right gap points have speeds ||v_l|| and ||v_r||, which generally differ from v_g = ||κ v_l + (1−κ) v_r||; their PN headings require K_l = v_e/||v_l|| and K_r = v_e/||v_r||, not K. Furthermore, even if an interval bound on θ_e were valid, the robot's actual velocity direction is γ_e = θ_e + β_g, whereas the endpoint directions are γ_{e/r} = θ_{e/r} + β_r and γ_{e/l} = θ_{e/l} + β_l. Since β_g generally differs from β_r and β_l, an interval on θ does not place γ_e between the endpoint velocity directions. The final conclusion \"therefore performing collision-free gap passage\" is asserted rather than derived. The theorem may be repairable—for instance by directly analyzing the convex cone swept by the two gap poin","section":"§5.5.3, Eqs. (18)–(22)"},{"comment":"The proof assumes a \"feasible manipulated gap\" but Section 5.5.2 defines feasibility only as the existence of a PN intercept of the gap goal point. The proof jumps from \"the ego-robot will intercept the gap goal point between the left and right gap points\" to \"collision-free gap passage.\" This does not show that the entire trajectory up to that intercept remains inside the moving gap, nor does it account explicitly for the finite robot radius and the time at which the gap might close. The gap points are moving obstacles, and the gap tube model in Section 5.5.1 can include closure and re-opening; the theorem needs a precise statement of what \"passage\" means and why the straight-line PN trajectory is contained in the free space throughout the interval, not merely at the terminal intercept time.","section":"§5.5.3 and §5.5.2"},{"comment":"The Monte Carlo experiment is presented as demonstrating \"provably safe trajectories\" under ideal conditions, but it cannot substitute for the proof. Of 10,000 trials, 2,668 were declared kinematically infeasible and 345 timed out; only 6,987 executed a gap passage, and the sampling distribution over gap geometries is limited. The experiment is a reasonable sanity check that is consistent with the theorem, but the paper should not imply that it validates the universal claim. Once the proof of Theorem 1 is repaired, this experiment can serve as supporting evidence.","section":"§6.1"}],"minor_comments":[{"comment":"The statement that \"arctan is a monotonically increasing function\" is not by itself a sufficient argument for the interval inclusion of a convex combination of two-dimensional vectors. The authors should state the angular-span condition (convex polar triangle) and use the positive cone property explicitly.","section":"§5.5.3, Eqs. (19)–(20)"},{"comment":"The captions for panels (b) and (c) both say \"control thread,\" but the text indicates that one should be the planning thread. Please correct the labels.","section":"Figure 18"},{"comment":"The header \"Reported Computation Times (Hz)\" mixes units and quantity; \"planning rates\" would be clearer.","section":"§6.4.2, Table 1"},{"comment":"The social compliance cost formula is typeset in a garbled way; the vector notation and the definition of V_rel should be made explicit.","section":"§6.5, Eq. (26)"},{"comment":"The text refers to \"Figure 4.3\" for the information flow; the actual figure is numbered Figure 4. Please reconcile the cross-reference.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The practical planner, open-source release, and benchmarking effort are valuable, and the paper is generally well written. The main blocker is the proof of Theorem 1: the logical gap at the interval-to-heading step is central and cannot be fixed by the Monte Carlo experiment alone. The authors should either provide a rigorous proof (possibly using the convex-cone structure of the gap) or explicitly weaken the claim to an empirical statement. I would not reject the paper, because the gap appears repairable, but the formal claim as written overstates what is proven."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a real piece of work, not a desk reject. The planner itself is a coherent extension of the authors' own dynamic gap line, and it adds two things that matter: gap tubes that survive interruptions (closing and reopening), and ungap planning that lets the robot use receding obstacle space when no gap exists. The simulation benchmark against a broad set of classical, learned, and gap-based baselines is a genuine effort, the Arena-Rosnav integration and open-sourced code are concrete deliverables, and the hardware runs on a TurtleBot2 give the paper practical credibility. The gap propagation and scan-propagation algorithms are clearly specified and the timing analysis is useful.\n\nThe soft spots are real but not fatal. The strongest concern is the proof of Theorem 1. The stress-test note is right: Equation 22 applies a single speed ratio K=v_e/v_g to the left and right gap points, but those points have their own speeds, v_l and v_r, which generally differ from v_g. The intercept headings for the endpoints need K_l and K_r, not K. And even if the theta interval inference were patched, you still need to show the robot's actual velocity direction gamma_e = theta_e + beta_g lies between the endpoint directions; the current argument just asserts it. The theorem may be repairable — convexity of the convex combination is a plausible route — but as written, the formal guarantee is not established. The paper itself admits in Section 5.5.3 that few real gaps satisfy the constant-velocity and isolated-environment assumptions, so the practical weight falls on the propagation and safety layers, which are reasonable but not formally guaranteed.\n\nSecond, the abstract's claim that dynamic gap 'outperforms all other baselines in all environments' only holds for the holonomic version. The nonholonomic version lands at 36% success in the factory environment, and the paper's own conclusion acknowledges this. That is an overstatement that should be fixed.\n\nThe Monte Carlo experiment is a self-consistency check, not independent validation, but the paper doesn't pretend otherwise. The lack of error bars on the 25-trial benchmarks is a minor annoyance, not a fatal flaw.\n\nBottom line: this deserves peer review. The engineering is solid, the ideas extend the subfield, and the central proof gap is identifiable and potentially fixable. I'd want a referee to push on the theorem before acceptance, but this is exactly the kind of paper a serious venue should engage with, not desk reject.","headline":"A serious, well-engineered extension of gap-based planning that ships code and strong benchmarks, but the central formal claim in Theorem 1 is not actually proven as written and the abstract oversells the results.","tokens_in":27178,"tokens_out":901,"would_cite":false,"duration_ms":12913,"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 gap-based planner can steer a robot through moving free-space gaps with a formal collision-free guarantee under ideal conditions.","keywords":["gap-based planning","dynamic obstacle avoidance","Parallel Navigation","pursuit guidance","free-space tracking","mobile robot navigation","collision-free guarantee","gap propagation"],"falsifier":"Run the PN policy through a single manipulated gap whose endpoints are genuinely constant-velocity (measured, not assumed) in an otherwise empty environment; a collision before intercept would disprove Theorem 1. Conversely, a corridor test with an accelerating pedestrian, where the planner nevertheless commits to a gap tube, would show the constant-velocity condition is doing the safety work.","tokens_in":26293,"feed_emoji":"🤖","tokens_out":5201,"duration_ms":61391,"temperature":0.7,"pith_summary":"The paper claims that a robot can navigate dynamic environments by treating free space as moving geometric objects—gaps—and predicting how those gaps evolve, rather than reacting to obstacles frame by frame. It introduces a planner that detects gaps in laser scans, tracks their endpoints across time with a Kalman filter, propagates them forward to build 'gap tubes' describing future availability, and then applies the Parallel Navigation guidance law from pursuit theory to steer toward a point inside a feasible gap. Under ideal conditions—a first-order holonomic robot, constant-velocity gap endpoints, and an isolated local environment—the paper proves the PN policy yields collision-free passage through a moving gap. It also handles the opposite case, planning through occupied polar regions ('ungaps') when no gap exists. If the claims hold, the result is a local planner whose collision avoidance rests on a formal guarantee rather than emergent robustness, and the paper reports that it outperformed classical and learned baselines in simulation and ran successfully on real hardware.","feed_headline":"Robot passes moving gaps with a provable safety guarantee","feed_subtitle":"Tracking free-space gaps and steering with Parallel Navigation beats eight baselines in simulation and runs on a real robot.","key_machinery":"The enabling objects are gap tubes and the Parallel Navigation (PN) geometric rule. A gap is a polar wedge of free space between two moving endpoints (left and right gap points); a gap tube is a temporal sequence of gaps and lifespans that records how a gap closes, reopens, or changes dynamics over the planning horizon. PN, the constant-bearing guidance law, supplies the kinematics: the robot steers so the bearing to the gap goal stays fixed while range decreases, which gives explicit conditions for interception and an intercept time. Added machinery includes gap detection and simplification from raw scans, Hungarian-assignment tracking of gap points, a constant-velocity extended Kalman filt","core_discovery":"The central claim is Theorem 1: for a first-order holonomic ego-robot, a feasible manipulated gap whose left and right points move at constant velocity, and an isolated environment, steering with the Parallel Navigation (constant-bearing) policy toward the gap goal point p_g yields collision-free passage. The proof works by defining the gap goal position and velocity as convex combinations of the two endpoint states, then observing that bearings and intercept headings are order-preserving under convex combination and monotone trigonometric functions. Because the ego-robot's intercept heading lies between the headings that would intercept the left and right endpoints, it passes between them.","pith_inferences":["A natural stress test the paper does not run: replace the constant-velocity assumption with bounded acceleration, and check whether inflating the gap or widening the feasibility inequalities preserves collision-free passage; the tube representation would likely need a safety margin that grows with the acceleration bound.","The proof's convexity argument suggests the guarantee is not specific to PN: any intercept law whose commanded bearing varies monotonically with the gap goal bearing should preserve the between-the-endpoints property, so the same framework could host other guidance laws.","The gap-tube representation could be coupled with learned trajectory predictors: instead of assuming constant velocity, plug predicted endpoint distributions into the propagation step and score gap tubes by probability of remaining open; this would extend the formal core to uncertain, human-dominated scenes without abandoning it.","The model mismatch noted for the nonholonomic robot is a concrete next target: integrating turning constraints into the PN feasibility check, rather than planning holonomic and tracking with a nonholonomic controller, may close the performance gap."],"forward_implications":["Gap-based planners can be made explicitly dynamic and perception-informed: free space is tracked as a stateful object rather than re-detected from scratch each scan.","Trajectory selection acquires a formal safety anchor: feasible gap tubes come with an intercept-time condition, so the planner can refuse gaps that will close before the robot arrives.","The planner can exploit gaps that are currently closed but predicted to reopen, and can follow receding occupied regions ('ungaps') when no gap is passable.","In simulation benchmarks across empty, factory, and hospital environments with fifteen dynamic agents, the holonomic version of dynamic gap outperformed all classical and learned baselines, while the nonholonomic variant performed worse due to model mismatch.","Hardware tests on a differential-drive robot confirmed the planner's behaviors in real dynamic environments."],"supporting_citations":[{"why":"Supplies the static gap-based predecessor, including gap manipulation, trajectory scoring, and the projection-operator safety filter that dynamic gap inherits.","marker":"[2]"},{"why":"The authors' earlier proof-of-concept dynamic gap planner, whose isolated-gap and unsafe-ungap assumptions this chapter relaxes.","marker":"[7]"},{"why":"Prior work extending gap-based planners to dynamic environments, identified as relying on ground-truth agent states rather than perception-informed sensing.","marker":"[5]"},{"why":"Predictive follow-the-gap method for dynamic obstacles; a prior dynamic gap-based planner used as comparison for the perception-informed contribution.","marker":"[6]"},{"why":"Source of the Pure Pursuit and Parallel Navigation guidance-law kinematics used in feasibility analysis and the proof of Theorem 1.","marker":"[81]"},{"why":"Hungarian assignment method used for associating gap points and propagated gaps across timesteps.","marker":"[20]"},{"why":"Benchmarking platform in which the proposed planner and baselines are evaluated in simulation.","marker":"[8]"},{"why":"Introduces planning in the perception space, the paradigm that motivates keeping laser scans in egocentric polar form.","marker":"[65]"}],"fun_headline_variants":["Gap-tracking planner dodges moving obstacles with proof","Provably safe robot navigation through dynamic free-space gaps","Dynamic-gap planner beats 8 baselines in moving scenes","Safe motion planning by tracking and propagating gaps in time"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The planner's predictions and its safety proof both assume the gap's left and right points move at constant velocity over the planning horizon; if the agents accelerate, turn, or change speed, the propagated gap tubes are no longer correct and the collision-free guarantee does not apply.","fun_headline_variants_meta":{"raw":{"variants":["Gap-tracking planner dodges moving obstacles with proof","Provably safe robot navigation through dynamic free-space gaps","Dynamic-gap planner beats 8 baselines in moving scenes","Safe motion planning by tracking and propagating gaps in time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":8.8e-05,"raw_usage":{"total_tokens":788,"prompt_tokens":698,"completion_tokens":90,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":24}},"tokens_in":442,"tokens_out":90,"duration_ms":2017,"temperature":1.0,"reasoning_tokens":24,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:34:42.883546+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the PN policy through a single manipulated gap whose endpoints are genuinely constant-velocity (measured, not assumed) in an otherwise empty environment; a collision before intercept would disprove Theorem 1. Conversely, a corridor test with an accelerating pedestrian, where the planner nevertheless commits to a gap tube, would show the constant-velocity condition is doing the safety work.","supporting_citations":[{"cited_title":"Shneydor,Missile Guidance and Pursuit: Kinematics, Dynamics and Control (1st ed.)","cited_arxiv_id":null,"evidence_quote":"Source of the Pure Pursuit and Parallel Navigation guidance-law kinematics used in feasibility analysis and the proof of Theorem 1."},{"cited_title":"The Hungarian method for the assignment problem,","cited_arxiv_id":null,"evidence_quote":"Hungarian assignment method used for associating gap points and propagated gaps across timesteps."},{"cited_title":"PiPS: Planning in perception space,","cited_arxiv_id":null,"evidence_quote":"Introduces planning in the perception space, the paradigm that motivates keeping laser scans in egocentric polar form."}],"review_version":1}