{"id":"95824d35-e639-44f3-82ed-1bc8d5c92a84","arxiv_id":"2607.24369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Mixed-integer tracking MPC with a shortest-path offset cost plans dynamically feasible UAV trajectories in non-convex zonotopic air corridors and reaches the target without an external global planner.","lead":"A mixed-integer tracking MPC planner steers UAVs through non-convex urban air corridors by embedding a shortest-path offset cost so the vehicle converges without a separate global planner. It matters for Urban Air Mobility, where corridors are non-convex and real-time, dynamically feasible references are required.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The convergence mechanism rests on the offset path being a genuinely feasible shortest path, but containment is only checked at finitely many interpolated points (Np=2), so segments of the \"shortest feasible path\" can silently exit the corridor — undercutting the abstract's \"guarantee convergence\"措辞","rationale":"The reader's weakest_assumption identifies exactly this load-bearing point: finite-point zonotopic containment (Np=2) is insufficient to certify continuous-segment feasibility of the artificial path, and continuous verification is deferred to future work. My pass confirms this is the right soft spot and sharpens why it matters: it is not just a safety/verification nicety, it is the premise of the offset-cost convergence argument, since the offset must measure a genuinely feasible path length for the global-convergence intuition from Köhler et al. (2025) to transfer. I considered competing concerns — (i) no recursive-feasibility or convergence theorem for the MI problem, (ii) a single deterministic simulation with no baselines, (iii) the ns=nz sufficiency claim in Fig. 3 — but (i) and (ii) are already priced into the reader's CONDITIONAL verdict and MODERATE confidence, and (iii) is plausible for chain-structured corridors and secondary. I also checked the §3.1 claim that ys∈Z follows implicitly from the terminal equality x(N)=xs; that reasoning is sound since y(N)∈Z is enforced by (4). The reader's verdict already conditions on softening the guarantee language and adding reproducibility artifacts, which is precisely what this concern demands; hence UNCHANGED rather than a downgrade. The proposed test is cheap (post-hoc continuous verification of one trajectory, plus one re-solve at higher Np) and would settle whether the concern is live for the reported scenario or only for the general claim.","tokens_in":9577,"tokens_out":2309,"duration_ms":85194,"concrete_test":"Take the optimized artificial path r from the §5 scenario and verify continuous containment: for each segment q, solve max over t∈[0,1] of dist(r(q)+t(r(q+1)−r(q)), Z) using the H-representation of the four zonotopes (or check intersection with the complement). If any segment exits Z, Vo underestimates the feasible shortest path and the convergence mechanism is broken as stated. Then re-solve problem (9) with Np=8 and with an exact segment-containment constraint; if the optimized offset cost or trajectory changes materially, the Np=2 result was an artifact of under-checking.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that the MI tracking MPC with the embedded shortest-path offset \"converges to the target\" in the non-convex corridor. The mechanism that is supposed to deliver this — inherited from Soloperto et al. (2023a) / Köhler et al. (2025) — requires the offset cost Vo(r) in (7) to be (a proxy for) the length of a genuinely feasible path from ys to yt through Z. In problem (8)/(9), feasibility of each artificial-path segment is enforced only at the discrete interpolation set A = {k/Np}, via (5)–(6) and (8e)/(9l). Nothing in the formulation constrains the straight line between two consecutive checked points: with Np=2 (the experiment's setting), only the segment endpoints and midpoint are verified, and ζ_i(q,α) is free to select a different zonotope at each α. A straight segment between a midpoint in Z_i and an endpoint in Z_j (i≠j) can pass outside Z=∪Z_i between checked points — e.g., cutting the outside corner at an L-bend of the orthogonal corridor used in §5. In that case Vo(r) strictly underestimates the true shortest feasible path length, the offset no longer upper-bounds/geodesically measures distance to the target, and the Lyapunov-style convergence argument the paper implicitly relies on (\"By minimizing this piecewise-linear path, the offset cost induces global convergence,\" §3.2) loses its premise. This is a correctness gap in the guarantee, not merely a consensus difference: the paper claims a guarantee while deferring \"geometric methods for continuous segment verification\" to future work (§6). Note the same finite-check pattern applies to the actual UAV trajectory only at discrete time steps (τs=0.5 s), which the authors at least acknowledge as tunneling risk; the artificial-path issue is unacknowledged and load-bearing for the headline claim. A mitigating observation: for this specific test geometry (axis-aligned rectangular zonotopes joined at full faces), midpoint-plus-endpoint checking may happen to suffice, which is presumably why the s","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes a motion planner for UAV navigation in non-convex urban air corridors, formulated as a single mixed-integer tracking MPC (problem (9)). Corridor membership is enforced via Big-M zonotope containment with binary segment assignment (Eqs. (4)–(6)), and convergence in the non-convex domain is sought by replacing the Euclidean offset cost of standard tracking MPC with a shortest-path-inspired offset Vo(r) (Eq. (7)) over an ns-segment artificial path, following Köhler et al. (2025). A geometry-based warm start (problem (10)) restricts intermediate path nodes to consecutive zonotope intersections, removing the external roadmap required by the parent method. A single numerical study (four orthogonal rectangular zonotopes, point-mass dynamics (11), N=5, τs=0.5 s, Np=2) shows a constraint-compliant trajectory reaching the target.","tokens_in":10011,"tokens_out":3438,"duration_ms":118319,"significance":"If the formulation performs as claimed, it is a useful contribution: a single MI optimization that simultaneously enforces dynamic feasibility, corridor containment, and global guidance, without a separate geometric planner. The warm-start construction (§4.1, problem (10)) is a genuinely practical element, and the manuscript is honest about several limitations (tunneling at τs=0.5 s, solver scaling, reliance on Bonmin). The claimed payoff, however, is precisely the convergence guarantee in the Abstract and §6, and that is where the manuscript is weakest: the guarantee is asserted, not established, and the mechanism that is supposed to deliver it has a concrete correctness gap (finite-point containment, see Major Comment 1). The result is currently a promising heuristic planner with one hand-tuned demonstration, not a planner with a guaranteed property. The gap appears fixable within the paper's scope, hence the recommendation.","major_comments":[{"comment":"§3.2, Eqs. (5)–(6), (8c)–(8e), (9h)–(9i), (9l): containment of the artificial path is enforced only at the finite interpolation set A={k/Np}, and the segment-assignment binaries ζi(q,α) are free to select a different zonotope at each α. With the experiment's Np=2, only endpoints and the midpoint of each straight segment are checked; the open sub-intervals between checked points are unconstrained, so a segment can exit Z=∪Zi (e.g., cut the outside corner at the L-bend of the §5 corridor) while satisfying all constraints. In that case Vo(r) (7) strictly underestimates the true shortest feasible path length, and the Lyapunov/geodesic argument inherited from Soloperto et al. (2023a)/Köhler et al. (2025) — invoked in §3.2 ('the offset cost induces global convergence') and behind the Abstract's 'guarantee convergence' — loses its premise. The authors appear aware (§6 defers 'geometric methods","section":"§3.2 / Eqs. (5)–(8), (9h)–(9l)"},{"comment":"Abstract, §3.2, §6: no convergence statement is actually proved. There is no theorem, proposition, or lemma establishing recursive feasibility and convergence of the receding-horizon scheme (9) in the non-convex setting — the text asserts the conclusion ('ensures global target convergence', §6) by analogy with Köhler et al. (2025). Even granting continuous containment, transferring that result requires verifying its hypotheses here: the relationship between the offset weight κs (ks=50) and the stage-cost weights (Q,R), the role of the terminal equality x(N)=xs with a short horizon N=5, and the shrinkage λ in (2). The manuscript should either (i) provide a formal statement with proof (or an explicit, hypothesis-by-hypothesis reduction to the cited result), or (ii) downgrade the claim to 'mitigates local minima / converges in simulation' and frame the guarantee as conjectural. As written,","section":"Abstract / §3.2 / §6"},{"comment":"Contribution (iii) is 'a geometry-based warm start for real-time operation' (§1, §4.1), yet the paper reports no computation times at all. §5 names the toolchain (Matlab 2025a, CasADi, Bonmin) but gives neither solve times per MPC step, nor a comparison with/without the warm start (problem (10)), nor the number of binary variables in the instance (nz=4 zonotopes, N=5, Np=2, ns=4 implies a specific count that should be stated). Additionally, Bonmin is a local/heuristic MINLP solver: it does not certify global optimality of the mixed-integer subproblem, so even a correct offset-cost design could silently return a suboptimal artificial path in harder topologies. A minimal table of per-step solve times (with and without warm start, plus one larger-nz case) and a sentence on solver choice would substantially support the real-time claim.","section":"§4.1 / §5"},{"comment":"§5: validation rests on a single hand-tuned scenario with one target and one initial condition. The failure mode the method is designed to avoid — the Euclidean-offset local minimum of Fig. 2(a) — is never simulated: there is no baseline run of the standard tracking MPC (Euclidean Vo) in the same corridor, so the reader cannot verify that the proposed offset is doing the claimed work in this instance. A baseline comparison, and at least one additional configuration (e.g., a target behind two bends, or a U-shaped corridor where the shortcut-through-the-wall failure is sharpest), would materially strengthen the evidence for the central claim.","section":"§5 / Fig. 2"}],"minor_comments":[{"comment":"Cross-reference errors: §4, second paragraph refers to 'the optimization problem in (19)' — should be (9). Fig. 5 caption cites 'the original problem (18)' and 'the auxiliary optimization problem (20)' — should be (8) and (10). §3.1 refers to 'the predicted output feasibility constraints of (9)' before problem (9) is introduced (presumably (3) is meant).","section":"§3.1 / §4 / Fig. 5"},{"comment":"Typos/OCR artifacts: 'UA V', 'F AA', and 'K¨ ohler' (with spurious space) recur throughout, including the Abstract; '(x(k), u(k))∈X×U' in §2 is missing its opening parenthesis. Inconsistent index ranges: (3c) uses j∈N0:N while (9c) uses N0:N−1 for the same operational constraint.","section":"Throughout"},{"comment":"§5 states 'two interpolation points (Np=2)', but per the definition A={k/Np | k∈N0:Np} this yields three evaluated points (α=0, 1/2, 1) per segment. Please reconcile the wording; the actual binary-variable count depends on it.","section":"§5"},{"comment":"§3.1: the argument that ys∈Z needs no explicit constraint because x(N)=xs implies ys=h(x(N)) deserves one more sentence: it requires that the terminal output actually satisfies the corridor constraint at j=N and that h is the same map used in (3d). As (9c) excludes j=N from the operational constraints, please confirm (4)/(9d) is still imposed at j=N (the index range in (9d) says N0:N, which is consistent — but (9c) is not).","section":"§3.1 / Eq. (9)"},{"comment":"§4.1, Fig. 3: the claim that 'ns=nz is always sufficient to represent the shortest path' is stated without justification. It is plausible for a linearly ordered chain of pairwise-border-intersecting zonotopes, but for general unions of nz convex sets a shortest feasible path can require revisiting a region or bending multiple times at one intersection. Please state the precise corridor-topology assumption under which the claim holds.","section":"§4.1 / Fig. 3"},{"comment":"Figures 3–5 are referenced as analytical aids but are described only in captions; a brief in-text walkthrough of Fig. 4(a)–(d) (what changes between panels) would help. Fig. 6–9: please add axis labels/units and mark the zonotope boundaries and target in Fig. 6(b) for readability.","section":"§4.1 / §5 figures"},{"comment":"Köhler et al. (2025) is cited as arXiv:2509.15917 — an unpublished preprint. Since the convergence mechanism is imported wholesale from it, the paper's claims are partly hostage to that reference's status; at minimum this dependency should be flagged in the text.","section":"References / §3.2"}],"recommendation":"major_revision","confidential_remarks":"The formulation is a competent adaptation of Köhler et al. (2025) to corridor geometry, with the warm start (10) as the clearest novel element; novelty is incremental but within scope. The decisive issue is the gap between the advertised 'guarantee' and what is established: finite-point containment (Np=2) plus an unproved convergence assertion. Both look repairable — a per-segment (rather than per-interpolation-point) zonotope assignment likely restores continuous containment by convexity, and the guarantee language can otherwise be softened — but this requires real work, not editing. Citation pattern is unremarkable; self-citations supply background, not results. If the venue is an IFAC-style conference with a hard page limit, the authors may prefer option (ii) of Major Comment 2 (qualified claims); for a journal version I would push for the formal statement."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent domain adaptation, not a new theory. They fold Limon-style tracking MPC, Big-M zonotope assignment, and a Köhler/Soloperto-style shortest-path offset into one mixed-integer program aimed at structured air corridors, plus a warm start that pins intermediate path nodes to consecutive zonotope intersections so you do not need an external roadmap. That warm-start and the single-shot formulation are the real artifacts.\n\nWhat works: formulation (9) is readable and coherent. The corridor geometry is used honestly for choosing ns = nz and for the restricted init problem (10). The single four-zonotope orthogonal-canyon run produces a dynamically feasible, corridor-respecting trajectory that reaches the target with N=5, which is what they claim the architecture can do. Citations are appropriate; self-cites are background, not circular. Circularity burden is low—this is a methods paper that designs a cost and shows a sim.\n\nSoft spots, in proportion: the abstract and intro say “guarantee convergence.” There is no theorem for the non-convex MI case. The offset path only checks containment at finitely many interpolated points (Np=2 in the experiment). Straight segments between checked points can leave the union of zonotopes, especially at L-bends, so Vo(r) can understate true feasible length and the geodesic argument loses its premise. They defer continuous-segment verification to future work while still using guarantee language—that is the load-bearing gap. Empirics are one hand-tuned deterministic scenario, no baselines, no timing tables worth much, no code. Tunneling on the actual trajectory is at least acknowledged; the artificial-path issue is not. Free parameters (N, τs, κs, Q/R, M, Np, ns, λ) are many and scenario-specific.\n\nWho it is for: people already building corridor MPC for UAM who want a single-layer alternative to geometric planner + tracker. Not a general non-convex MPC advance. Math is checkable from the text; data are thin but not fabricated.\n\nI would send it to peer review at a solid robotics/control venue (workshop or conference track). Soften the guarantee claims, add at least one baseline or ablation and a second geometry, and require continuous-feasibility discussion or safety margins. Worth engaging if you work this niche; skip if you need theory or multi-scenario evidence.","headline":"Clean packaging of known tracking-MPC + shortest-path offset ideas into one MI planner for UAM corridors; useful engineering, overstated “guarantee,” thin empirics.","tokens_in":11007,"tokens_out":600,"would_cite":false,"duration_ms":17943,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single mixed-integer tracking MPC with a shortest-path offset cost plans dynamically feasible UAV paths through non-convex air corridors and reaches the target without a separate global planner.","keywords":["Optimal Motion Planning","MPC","Urban Air Mobility","Unmanned Aerial Vehicles","non-convex corridors","tracking MPC","zonotopes","mixed-integer programming"],"falsifier":"Run the same four-zonotope canyon scenario with Np = 2 and show either that the continuous straight-line segments between nodes leave the corridor, or that the closed-loop trajectory still stalls in a local minimum short of the target; either outcome would break the claimed guarantee.","tokens_in":10606,"feed_emoji":"🛸","tokens_out":819,"duration_ms":19238,"temperature":0.7,"pith_summary":"Urban air corridors are non-convex unions of convex segments, so ordinary tracking controllers can trap a UAV in a local minimum far from the goal. This paper embeds corridor membership and a shortest feasible path from an artificial reference to the true target inside one mixed-integer model-predictive planning problem. The optimizer therefore produces a short-horizon trajectory that stays inside the corridor, respects vehicle dynamics and actuator limits, and steadily drives the artificial reference toward the destination. Simulations in a four-segment urban-canyon corridor show smooth, constraint-satisfying flight from start to goal with only a five-step horizon and no external roadmap. The practical payoff is a planner that can re-plan online when the target or the corridor changes, while still handing dynamically valid references to a low-level stabilizer.","feed_headline":"One MPC steers UAVs through non-convex air corridors","feed_subtitle":"A shortest-path offset cost replaces the global planner and still reaches the target","key_machinery":"The shortest-path offset cost Vo(r): a weighted length of an ns-segment path from the artificial output ys to the target yt, with every interpolated point forced inside exactly one zonotopic corridor segment by binary indicators and Big-M constraints; this cost replaces the usual Euclidean offset and is solved inside the same MIP as the dynamics.","core_discovery":"A mixed-integer tracking MPC that jointly optimizes the predicted trajectory, an artificial steady-state, and a piecewise-linear shortest path connecting that artificial output to the true target generates dynamically consistent, corridor-feasible references that converge to the goal in non-convex airspace without any external global planning stage.","pith_inferences":["If continuous-segment verification (via hybrid zonotopes) succeeds, the same offset-cost idea could extend to free-flight regions that are not pre-partitioned into corridors.","The aggressive force profiles seen in simulation suggest that adding input-rate penalties would trade a little path length for smoother actuator use without losing the convergence property.","Because the warm-start already exploits consecutive zonotope intersections, the method may scale better to longer corridor networks than a generic mixed-integer shortest-path solver."],"forward_implications":["UAV motion planning in structured UAM corridors can drop the separate geometric global planner and still reach changing targets.","Corridor geometry alone (ns = nz and intersection-based warm-start) supplies a feasible first guess, removing the need for a precomputed roadmap.","Short prediction horizons remain sufficient once the offset cost encodes global connectivity.","The same MIP structure can later absorb multi-agent constraints or reactive obstacles once integer complexity is reduced."],"fun_headline_variants":["Mixed-integer MPC plans UAV paths in non-convex corridors","One MPC tracks shortest-path offsets through non-convex airspace","MPC embeds shortest-path cost for corridor-feasible UAV flight","Mixed-integer planner steers UAVs without external global stage","Tracking MPC converges UAVs in non-convex urban air corridors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Checking that only a few discrete sample points along each artificial-path segment lie inside the corridor is enough to keep the whole continuous path inside and thereby guarantee global convergence.","fun_headline_variants_meta":{"raw":{"variants":["Mixed-integer MPC plans UAV paths in non-convex corridors","One MPC tracks shortest-path offsets through non-convex airspace","MPC embeds shortest-path cost for corridor-feasible UAV flight","Mixed-integer planner steers UAVs without external global stage","Tracking MPC converges UAVs in non-convex urban air corridors"]},"model":"grok-4.5","effort":"low","cost_usd":0.002152,"raw_usage":{"total_tokens":807,"prompt_tokens":618,"num_sources_used":0,"completion_tokens":74,"cost_in_usd_ticks":21524000,"prompt_tokens_details":{"text_tokens":618,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":115,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":618,"tokens_out":74,"duration_ms":3394,"temperature":1.0,"reasoning_tokens":115,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T16:41:01.705800+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same four-zonotope canyon scenario with Np = 2 and show either that the continuous straight-line segments between nodes leave the corridor, or that the closed-loop trajectory still stalls in a local minimum short of the target; either outcome would break the claimed guarantee.","supporting_citations":[],"review_version":1}