{"id":"147f19b6-ed15-4406-bab4-02c651c130ca","arxiv_id":"2507.21772","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An edge-minimizing RRT* tree planner combined with distributed MPC generates and maintains line-of-sight-connected, collision-free UAV relay deployments in cluttered 3D spaces.","lead":"This paper describes a two-step system that first plans a tree-shaped relay network for drones in an obstacle-filled area, then uses distributed model predictive control to fly the drones along the network while keeping wireless links clear of obstacles. It aims to reduce the number of relay drones needed and to guarantee collision-free, connected motion, with simulations and small-drone experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's recursive-feasibility proof is asserted rather than shown: the shifted previous plan may violate the recomputed terminal-step connectivity constraints because the (K+1)-th point is updated online.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: Theorem 1 depends on a recursive-feasibility invariant for constraints (6) and (9) that is asserted but not proven. My stress-test sharpens this concern by locating the specific terminal-step mechanism: constraints at k=K depend on p_i^{K+1}(t), which is updated online and therefore need not equal the previous plan's p_i^{K+1}(t-h). Hence the shifted previous trajectory is not automatically feasible for the recomputed constraints, and the proof's citation to Sec. III-B.2 does not resolve this. This is a correctness-risk issue with the paper's strongest theoretical claim, not a stylistic or incremental concern. However, it is potentially addressable by adding a genuine proof or by modifying the update rule for p_i^{K+1}; the simulations and hardware experiment provide supporting evidence but cannot substitute for the missing invariant. The reader's CONDITIONAL verdict is therefore appropriate: the architecture is plausible and likely fixable, but the headline guarantee should not be taken at face value until the recursive-feasibility lemma is actually established. I recommend no change to the reader's verdict, so verdict_should_be is UNCHANGED.","tokens_in":12787,"tokens_out":6435,"duration_ms":82856,"concrete_test":"Instrument the released code to log, at every MPC step t, the shifted predetermined trajectory P^i(t) = {p_i^2(t-h), ..., p_i^K(t-h), p_i^K(t-h)} and to evaluate whether it satisfies the newly recomputed constraints (6) and (9) for every edge in T, including step k=K with the online-updated p_i^{K+1}(t). Run the moving-target hardware scenario and a randomized set of obstacle/target instances. If any shifted plan violates a recomputed constraint, the proof's key assertion is false. If no violation is observed over many steps, the invariant is empirically supported but still lacks the analytic proof that Theorem 1 requires.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee (Theorem 1) rests entirely on recursive feasibility of the distributed MPC (12). The proof asserts that the shifted predetermined trajectory satisfies the newly recomputed constraints (6) and (9) because 'the constraints (6) and (9) are satisfied at t if they are satisfied at t-h, as shown in Sec. III-B.2.' But Sec. III-B.2 derives the constraints and never proves this persistence claim. The gap is not merely expository. At step k=K, constraints (6) and (9) are defined using the (K+1)-th point p_i^{K+1}(t)=p_i^{tg}(t), which is updated online for searchers. The shifted plan provides p_i^k(t)=p_i^{k+1}(t-h) for k<K and p_i^K(t)=p_i^K(t-h), but p_i^{K+1}(t) is generally different from p_i^{K+1}(t-h), especially when targets move online. Thus the center c_ij^K and the separating hyperplane in (8)/(9) are recomputed with new points, and no argument is given that the old p_i^K and p_j^K remain inside the new sphere or the new half-space. The proof's citation to Sec. III-B.2 does not supply that argument. Feasibility of the inter-agent and obstacle constraints (4) and (5) is also imported from prior works [24], [25] rather than re-derived for the coupled problem, but the terminal-step connectivity persistence issue is the load-bearing one. If this invariant fails, Theorem 1 collapses even if simulations and hardware experiments look successful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-layer approach for deploying a team of UAVs as a multi-hop relay network in obstacle-cluttered environments. The first layer (Alg. 1, MiniEdgeRRT*) constructs a spanning-tree topology to minimize the required number of relay agents. The second layer is a distributed MPC (Alg. 2, problem (12)) that encodes inter-agent collision avoidance via buffered Voronoi cells and safe corridors, and connectivity maintenance via distance-sphere constraints (6) and line-of-sight separating-hyperplane constraints (9). The paper claims a theoretical guarantee (Theorem 1) that the team remains collision-free and connected at all times, and a homotopy-preservation claim (Theorem 2). The approach is evaluated in 3D simulations and in hardware experiments with a moving target, and the code is made available.","tokens_in":13095,"tokens_out":8344,"duration_ms":97359,"significance":"If the theoretical guarantee were fully established, the paper would make a useful contribution: it treats LOS connectivity as a hard constraint in a distributed MPC with convex constraints, and it optimizes the number of relay agents rather than assuming a fixed team. The empirical evaluation is broad, including comparison with an MILP baseline, and the release of source code supports reproducibility. However, the central guarantee depends on a recursive-feasibility argument that is asserted rather than proved in the manuscript, and the proof also imports key feasibility results from prior self-cited work. These gaps prevent the paper from substantiating its main advertised claim as written.","major_comments":[{"comment":"The proof of Theorem 1 states that the optimization (12) is recursively feasible because \"constraints (6) and (9) are satisfied at t if they are satisfied at t-h, as shown in Sec. III-B.2\", but Sec. III-B.2 contains no such proof; it only defines the coefficients c_ij^k and the separating hyperplane coefficients and states the separating condition on the current P_free. The gap is load-bearing: for a searcher, p_i^{K+1}(t)=p_i^{tg}(t) is updated online, so at step k=K the center c_ij^K in (6) and the coefficients in (8)-(9) are recomputed from new points p_i^{K+1}(t) and p_j^{K+1}(t). The shifted previous plan provides p_i^K(t)=p_i^K(t-h) and p_i^{K+1}(t-h), but not p_i^{K+1}(t); hence the old point p_i^K(t) can lie outside the new sphere or the new half-space. Without a proof of this persistence property, or a modification of the algorithm (e.g., fixing p_i^{K+1} over the horizon or adding a terminal constraint that makes the shifted plan feasible for the recomputed constraints), Theorem 1 is not established. The fallback to the predetermined trajectory in Sec. III-C does not repair this, since that trajectory is not shown to satisfy the newly recomputed constraints.","section":"III-D, Theorem 1"},{"comment":"Theorem 1 states that the system remains collision-free and connected \"at all times,\" but the proof only establishes feasibility of the optimization (12) at successive sampling instants. The constraints (6) and (9) are enforced on the planned positions p_i^k at discrete steps k, while the actual continuous-time trajectory between samples follows the quadratic arc generated by the double-integrator model with a constant input. No margin analogous to r1_min in (4) is introduced for the communication-range constraint (6) or the LOS constraint (9), so the paper does not rule out a violation of LOS connectivity between sampling instants. The theorem's conclusion should be restricted to the sampling instants, or an inter-sample analysis should be supplied.","section":"III-D, III-B"},{"comment":"The proof of Theorem 1 delegates feasibility of constraints (4) and (5) to prior works [24] and [25] without verifying that the assumptions of those results hold here. In this paper, (4) and (5) are coupled with the online-updated intermediate target p_i^tg, with the terminal step K of the connectivity constraints, and with the fallback rule that selects the predetermined trajectory when (12) is infeasible. Since the fallback trajectory is not shown to satisfy the newly recomputed constraints (6) and (9), the fallback does not provide the claimed safety guarantee. The authors should either state the relevant feasibility results from [24] and [25] and prove that their hypotheses are satisfied in this coupled setting, or give a self-contained recursive-feasibility proof.","section":"III-D, III-B.1"}],"minor_comments":[{"comment":"The phrase \"the our method\" in the text around Table I is ungrammatical; also, reporting only the average over 10 runs without variance or confidence intervals makes the comparison with baselines less informative.","section":"IV-A.1, Table I"},{"comment":"Algorithm 2 has inconsistent notation: the input is denoted Ptg (the target set), but Line 1 calls ComNet(p_g, P_i^tg, O); the symbol P_i^tg is not defined there, and later p_i^tg denotes a single intermediate target. Please unify the notation.","section":"Alg. 2"},{"comment":"In case (ii) of the definition of c_ij^k, the condition \"if |p - c_i^k,center| < d_w for all p in {p_i^k, p_j^k, p_i^{k+1}, p_j^{k+1}}\" has a missing subscript on the center and an unusual use of the variable p; please clarify the condition.","section":"III-B.2"},{"comment":"The proof of Theorem 2 is explicitly a sketch and uses imprecise terms such as \"dynamic region\" and \"tractive point\"; if Theorem 2 is to be advertised as a theoretical property, the proof should be completed or the statement should be rephrased as an empirical observation.","section":"III-D, Theorem 2"},{"comment":"The comparison with the MILP method reports single-run results for a few scenarios; since MPC timing can depend strongly on the obstacle configuration, adding statistics over multiple runs would strengthen the claim of a 90% reduction.","section":"IV-A.2"},{"comment":"There are several formatting and typographical issues: the abstract and Section III-A.1 contain \"RRT ?\" with a stray symbol, and Remark 2 uses \"typologies\" instead of \"topologies\".","section":"Abstract, III-A.1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is that the advertised theoretical guarantee is not supported by the proof as written. I would recommend requesting a complete proof of the recursive-feasibility claim, or a revision of the claims to match what is actually proven. The paper also leans heavily on two self-citations ([24], [25]) for the feasibility of the collision-avoidance constraints; for a journal submission these results should be stated or reproduced in enough detail for the reviewer to verify. The empirical part is solid and the code availability is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something genuinely useful: it couples a spanning-tree relay planning method (an edge-minimizing RRT* variant) with a distributed MPC that encodes LOS connectivity as separating-hyperplane constraints, and it ships code plus a hardware demo. Turning the nonconvex LOS constraint into linear constraints inside a recursively feasible MPC is a real integration that I don't see in the prior work. If the guarantee holds, it's a practical subfield contribution.\n\nThe soft spot is exactly where the stress-test note lands. Theorem 1's proof asserts that constraints (6) and (9) persist from t-h to t, but Section III-B.2 doesn't prove that. For the terminal step, the (K+1)-th point is updated online (p_i^{K+1}=p_i^tg), so the sphere center and hyperplane get recomputed with new points. The shifted plan provides p_i^K(t)=p_i^K(t-h), but no argument shows that point stays inside the new sphere or half-space. That's load-bearing for the recursive-feasibility claim. The proof also imports feasibility of (4) and (5) from prior papers without re-deriving it here; that's acceptable as a citation, but the connectivity persistence is not covered by those citations.\n\nMinor issues: Table I reports no variance, and the text says \"less than 15% of relay agents\" which is consistent with the averages, but the VWMST baseline description is ambiguous. Theorem 2 is labeled a sketch, fine, but it's also not fully convincing.\n\nBottom line: the architecture is sensible, the simulations and hardware look reasonable, and the code is out. But the headline guarantee isn't supported as written. This is addressable: write a real inductive proof of the persistence claim, or state a weaker theorem. I'd send it to peer review and ask for that revision. The paper deserves referee time.","headline":"Useful integration of an edge-minimizing RRT* relay tree with distributed MPC enforcing LOS connectivity as linear constraints, but the Theorem 1 recursive-feasibility proof is asserted rather than shown.","tokens_in":13640,"tokens_out":2060,"would_cite":false,"duration_ms":21190,"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":"This paper claims that a two-layer planner—an RRT*-based topology search plus a distributed MPC with explicit distance-sphere and line-of-sight hyperplane constraints—can deploy a minimal team of UAVs in obstacle-cluttered 3D space while…","keywords":["multi-UAV deployment","line-of-sight connectivity","spanning tree","distributed model predictive control","collision avoidance","obstacle-cluttered environment","recursive feasibility"],"falsifier":"Log, over many replanning steps in a narrow or U-shaped obstacle course, whether the shifted predetermined trajectory P^i satisfies the freshly computed constraints (6) and (9); a single timestep where the shifted trajectory violates the new LOS or distance-sphere constraint while the system is still in the promised regime would refute the recursive-feasibility invariant on which Theorem 1 rests.","tokens_in":12553,"feed_emoji":"🚁","tokens_out":4727,"duration_ms":56367,"temperature":0.7,"pith_summary":"This paper aims to show that a UAV team can be deployed as a relay network in obstacle-cluttered 3D space without ever losing line-of-sight communication or crashing, and that the number of relay drones can be planned rather than assumed. The proposed solution first designs a spanning-tree communication topology using a modified RRT* that penalizes extra edges, then runs a distributed model predictive controller that encodes connectivity as explicit convex constraints. The paper's headline theoretical claim is Theorem 1: if the team starts collision-free and connected, it stays collision-free and the actual communication graph never loses the designed spanning tree. This matters because line-of-sight connectivity is normally a nonconvex, nonlinear condition that is approximated or solved slowly; here it is turned into linear inequalities solvable in tens of milliseconds.","feed_headline":"UAV relay fleets keep LOS links while dodging obstacles","feed_subtitle":"A distributed MPC with convex constraints promises collision-free, connected deployment using fewer relay drones.","key_machinery":"The central object is the embedded spanning tree T, a set of vertices (index, position, parent index) rooted at the ground station, together with a distributed MPC that encodes the tree's edges as convex constraints. Connectivity is split into a distance-sphere constraint (6) and a separating-hyperplane LOS constraint (9): for each edge, intermediate points are computed along the planned motion, and a hyperplane is fitted that separates the free-space convex hull from the inflated obstacle, making the nonconvex LOS condition a linear inequality. Collision avoidance uses a modified buffered Voronoi cell (4) and a convex safe corridor (5). The per-agent optimization (12) is a quadratically constrained quadratic program whose recursive feasibility is the load-bearing mechanism for the guarantee.","core_discovery":"The central claim is Theorem 1: given an initially collision-free and connected team, the distributed MPC keeps the system collision-free and the communication graph G(t) connected for all times, with the embedded spanning tree T always contained in G(t). The argument is recursive feasibility: the optimization at each step has a feasible solution obtained by shifting the previous planned trajectory one horizon forward, provided the distance-sphere constraints (6) and LOS hyperplane constraints (9) persist from one step to the next; the remaining collision-avoidance constraints are imported from the authors' earlier work. A second theorem states that the physical chain of agents from the ground station to each searcher remains topologically equivalent to the planned tree path, so the fleet does not just stay connected but keeps the intended relay structure.","pith_inferences":["A direct consequence the authors do not spell out is that the same hyperplane-separation idea could be applied to other nonconvex team constraints, such as sensor coverage or keep-out zones, as long as they can be separated from the previous plan's geometry.","The persistence claim behind Theorem 1 could be tested independently by logging, at each replan, whether the shifted previous trajectory violates the newly derived constraints; a violation would not necessarily invalidate the simulations but would mean the stated guarantee is stronger than what the proof supports.","Because the guarantee covers line-of-sight and range rather than the underlying radio channel, real-world packet loss may still occur even when the geometric constraints hold; extending the framework to signal-strength or ray-tracing models would be a natural next step."],"forward_implications":["The number of relay UAVs becomes part of the optimization, so missions can be sized to the minimum team rather than a redundant fleet.","The per-agent problem is a convex QCQP, which is why replanning takes tens of milliseconds instead of the seconds reported for MILP baselines.","The designed spanning tree remains a subgraph of the actual communication graph at all times, meaning the route from the ground station to every searcher is continuously available.","The dynamic-target hardware experiment shows the topology can adapt online when a target moves, not only during static deployment.","The method inherits the assumptions of its building blocks, including convex obstacles with known vertices and the buffered-Voronoi collision model."],"supporting_citations":[{"why":"Supplies the modified buffered Voronoi cell constraints and the earlier recursive-feasibility result for inter-agent collision avoidance, which the paper imports directly.","marker":"[24]"},{"why":"Supplies the convex safe-corridor obstacle constraints and their feasibility proof, which the paper relies on for the obstacle-avoidance part of the guarantee.","marker":"[25]"},{"why":"Provides the informed RRT* algorithm whose sampling and cost structure MiniEdgeRRT* modifies to minimize edge count.","marker":"[22]"},{"why":"Defines the visibility-weighted minimum spanning tree baseline and motivates line-of-sight-aware topology design.","marker":"[13]"},{"why":"Is the MILP-based LOS connectivity method used as the comparison baseline for computation time and trajectory quality.","marker":"[21]"},{"why":"Provides the distributed MPC formulation framework that the online motion coordination is built on.","marker":"[23]"}],"fun_headline_variants":["Distributed MPC keeps UAV fleet LOS-connected and safe","Fewer drones, guaranteed LOS: new MPC deployment","Relay UAVs: obstacle-safe with guaranteed LOS","Minimal relay fleet, full LOS connectivity via distributed MPC","Distributed MPC for multi-UAV: collision-free and LOS-connected"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee collapses if the shifted previous plan stops satisfying the newly recomputed distance-sphere and LOS hyperplane constraints, a persistence property the paper asserts but does not actually prove.","fun_headline_variants_meta":{"raw":{"variants":["Distributed MPC keeps UAV fleet LOS-connected and safe","Fewer drones, guaranteed LOS: new MPC deployment","Relay UAVs: obstacle-safe with guaranteed LOS","Minimal relay fleet, full LOS connectivity via distributed MPC","Distributed MPC for multi-UAV: collision-free and LOS-connected"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000984,"raw_usage":{"total_tokens":4154,"prompt_tokens":905,"completion_tokens":3249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":3169}},"tokens_in":521,"tokens_out":3249,"duration_ms":24967,"temperature":1.0,"reasoning_tokens":3169,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:23:49.779795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Log, over many replanning steps in a narrow or U-shaped obstacle course, whether the shifted predetermined trajectory P^i satisfies the freshly computed constraints (6) and (9); a single timestep where the shifted trajectory violates the new LOS or distance-sphere constraint while the system is still in the promised regime would refute the recursive-feasibility invariant on which Theorem 1 rests.","supporting_citations":[{"cited_title":"Deadlock resolution and recursive feasibility in mpc-based multi-robot trajectory generation,","cited_arxiv_id":null,"evidence_quote":"Supplies the modified buffered Voronoi cell constraints and the earlier recursive-feasibility result for inter-agent collision avoidance, which the paper imports directly."},{"cited_title":"Informed rrt*: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,","cited_arxiv_id":null,"evidence_quote":"Provides the informed RRT* algorithm whose sampling and cost structure MiniEdgeRRT* modifies to minimize edge count."},{"cited_title":"Graph connectivity control of a mobile robot network with mixed dynamic multi-tasks,","cited_arxiv_id":null,"evidence_quote":"Defines the visibility-weighted minimum spanning tree baseline and motivates line-of-sight-aware topology design."},{"cited_title":"Resilient robust connectivity for multiagent systems with line of sight using mixed-integer programming,","cited_arxiv_id":null,"evidence_quote":"Is the MILP-based LOS connectivity method used as the comparison baseline for computation time and trajectory quality."},{"cited_title":"A distributed model predictive control strategy for constrained multi-vehicle systems moving in unknown environments,","cited_arxiv_id":null,"evidence_quote":"Provides the distributed MPC formulation framework that the online motion coordination is built on."}],"review_version":1}