{"id":"59712245-2fe1-40f3-bd27-d63747757bca","arxiv_id":"2412.12437","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A simulation study that merges Lloyd's algorithm formation control with potential-field obstacle avoidance for multi-UAV swarms in 2D and 3D urban environments.","lead":"A team at the University of Oklahoma combines a fish-inspired formation algorithm with pigeon-inspired obstacle avoidance to steer simulated drone swarms through buildings and moving obstacles, including a 3D version. The paper is a simulation study, not a flight test, and the 3D maneuver is not supported by the paper's own equations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3D avoidance term is rotation-invariant: Eq. (20) equals (k_r/2)||p_i-o_k||^2, so Eq. (24) is not its gradient and no rotational steering mechanism exists.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: Eq. (24)'s gradient is analytically wrong because rotation matrices preserve norms, so the rotational potential in Eq. (20) collapses to a distance-only repulsion and the claimed 3D obstacle avoidance mechanism disappears. My independent check confirms this. This matters because the abstract and the stated contributions explicitly claim '3D obstacle avoidance maneuvers' as the novel extension beyond the prior 2D pigeon-inspired method; if the rotational potential supplies no directional force, the simulations in Case Study 3 do not exercise the mechanism the paper claims. I do not see a stronger concern: the Lloyd-based formation component is standard and plausible, the inter-vehicle collision term is a standard repulsion, and the absence of code, baselines, and multi-run statistics is a secondary issue compared to the mathematical invalidity of the central novelty. Therefore the reader's REJECT verdict stands without adjustment.","tokens_in":9877,"tokens_out":4786,"duration_ms":46273,"concrete_test":"Symbolically differentiate Eq. (20) with T_rx, T_ry, T_rz as defined in Eqs. (21)-(23), using either a computer algebra system or manual chain rule, and compare the resulting ∂U_r/∂p_i with Eq. (24). For constant alpha the correct gradient is k_r(p_i-o_k), not k_r T_rx T_ry T_rz(p_i-o_k); if alpha depends on distance, the result still differs from Eq. (24). A second, independent check: set all alpha-dependent rotations in Eqs. (22)-(23) to identity and rerun Case Study 3; if the interaction forces and trajectories are unchanged, the 3D maneuver term contributes nothing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing piece of the central claim is the extension to 3D obstacle avoidance in Section 2.4.2. In Eq. (20), U_r = (k_r/2)||T_rx T_ry T_rz(p_i-o_k)||^2. Since T_rx, T_ry, T_rz in Eqs. (21)-(23) are rotation matrices, their product is orthogonal and norm-preserving: U_r = (k_r/2)||p_i-o_k||^2 for every alpha. Thus the angle alpha from Eq. (16) cancels identically, and the potential adds only a distance-dependent radial repulsion, not the claimed 'additional maneuverability' or any 3D steering force. Eq. (24) states the gradient is k_r T_rx T_ry T_rz(||p_i-o_k||)∇(||p_i-o_k||), but differentiating Eq. (20) by the chain rule gives k_r(p_i-o_k) (plus, if alpha is treated as a function of position, extra terms that still cannot match Eq. (24)). Equivalently, the printed formula is a rotated radial vector field, which for a fixed rotation has nonzero curl and is not the gradient of any scalar potential. The planar term also has a related defect: differentiating ||k_v(p_i-o_k)|| involves k_v^T k_v(p_i-o_k)/||k_v(p_i-o_k)||, so Eq. (12) is not the correct gradient either. The Lloyd-based formation controller and the inter-vehicle repulsion term are not invalidated by this argument, but the central novelty—3D obstacle avoidance via rotational potentials—is an artifact of the incorrect gradient formula and does not exist in the model as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a hybrid control framework for multi-UAV formation flight: a centralized probabilistic Lloyd's algorithm determines optimal formation positions, a distributed Hooke's-law term handles inter-vehicle collision avoidance, and a pigeon-inspired potential field is used for obstacle avoidance, with a claimed extension to 3D via rotational potentials. The method is evaluated in three simulation case studies (8 and 12 UAVs) involving static and dynamic obstacles, buildings, formation changes, and formation recovery.","tokens_in":10327,"tokens_out":5609,"duration_ms":47118,"significance":"If correct, the framework would offer a relatively simple, non-consensus-based alternative for scalable multi-UAV formation control with 3D obstacle avoidance. However, the central mathematical contribution is not sound: the 3D rotational potential is invariant under the rotation matrices used, so it reduces to an ordinary radial repulsion, and the printed gradient formulas are not actual gradients. This collapses the claimed 3D maneuverability, which is the paper's primary novelty. The Lloyd-based formation controller and the inter-vehicle collision term may still be viable, but they are not the claimed contribution.","major_comments":[{"comment":"The rotational potential U_r in Eq. (20) is invariant under the rotation matrices T_rx, T_ry, T_rz defined in Eqs. (21)-(23), because the product of rotation matrices is orthogonal and thus preserves the Euclidean norm. Hence U_r(p_i,o_k) = (k_r/2)||p_i-o_k||^2 for every alpha, and the angle alpha in Eq. (16) cancels identically. Consequently, the claimed 'additional maneuverability' and the 3D steering force do not exist in the model as written. Furthermore, Eq. (24) is not the gradient of Eq. (20): differentiating Eq. (20) gives k_r(p_i-o_k) (with additional alpha-dependent terms if alpha is treated as position-dependent), whereas Eq. (24) is a rotated radial vector field that generally has nonzero curl and therefore cannot be the gradient of any scalar potential. This invalidates the central contribution of the paper.","section":"2.4.2 (Eqs. 20-24)"},{"comment":"The gradient of U_p given in Eq. (12) is incorrect for a non-identity diagonal matrix k_v. The chain rule applied to U_p = (1/2)(||k_v(p_i-o_k)|| - r_a)^2 yields ∇U_p = (||k_v(p_i-o_k)|| - r_a) * (k_v^T k_v(p_i-o_k)) / ||k_v(p_i-o_k)||, not the expression with ∇(||p_i-o_k||) as printed. As a consequence, the obstacle avoidance control law in Eq. (17) does not implement gradient descent on the declared potential, and the 'potential field' interpretation fails. Since Eq. (17) is also used in the 3D scenario, this error propagates into the main simulations.","section":"2.4.1 (Eqs. 11-12)"},{"comment":"The overall control law in Eq. (25) is formed by superposing the formation, collision, and obstacle avoidance terms, and the text claims stability from the 'superposition principle,' but no Lyapunov, invariance, or barrier-function argument is provided to show that the closed-loop system is collision-free in the presence of moving obstacles. The only evidence is the distance histories in Figures 5-7, which are qualitative and do not substantiate the 'collision-free' guarantee stated in the title and abstract.","section":"2.5 (Eq. 25)"}],"minor_comments":[{"comment":"Equation (2) is used twice with the same number: once for the velocity dynamics ḍ_i = u_i and once for the neighborhood definition N_i = {j : ||p_i-p_j|| < r_d}.","section":"Section 2 (Eqs. 1-2)"},{"comment":"In the sentence preceding Eq. (5), the cost function being differentiated is defined in Eq. (3), not Eq. (4); Eq. (4) defines the Voronoi partition.","section":"Section 2 (Eq. 5)"},{"comment":"The planar detection condition in Eq. (10) includes a term with z-coordinates (atan((o_kz-p_iz)/(o_kx-p_ix))), which is inconsistent with the stated planar 2D setting.","section":"Eq. (10)"},{"comment":"Table 2 lists r_ok in units of m/s, but r_ok is used as an obstacle radius (a length); also, the dynamic obstacle velocity in Case Study 2 (v_obs4 = [0.1,0.025,0]^T m/s) differs from the value in Table 2 ([0.2,0.05,0]).","section":"Table 2"},{"comment":"Reference [25] contains a typo in the title ('Dyamics' should be 'Dynamics'), and the text in Section 2.3 refers to a 'mass-spiring-damper system' rather than 'mass-spring-damper.'","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper presents itself as a 3D extension of the pigeon-inspired method in [3], but the extension is mathematically vacuous: the rotation matrices in the 3D potential cancel out because they preserve the norm, and the printed gradient formulas are not gradients. These are load-bearing errors in the central contribution. The simulations are self-made, lack comparison baselines, and do not provide quantitative success metrics (e.g., minimum inter-vehicle distance over many randomized runs). I recommend rejection. If the authors later introduce a genuinely orientation-dependent potential with correct gradients, a new submission could be considered, but the present manuscript's main claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper's central claim—a 3D extension of pigeon-inspired obstacle avoidance—does not survive contact with its own equations. The rotational potential in Eq. (20) is norm-invariant, so it reduces to the same radial repulsion as the planar term, and the claimed 3D maneuver is an artifact of an incorrect gradient.\n\nWhat's actually new: not much. The Lloyd-based formation controller with a moving barrier is a straightforward application of centroidal Voronoi tessellation, and the inter-vehicle collision law is taken from Guo et al. The paper does a decent job of assembling these pieces and demonstrating plausible swarm behavior in simulations with 8–12 UAVs. The writing is clear and the problem is relevant.\n\nThe soft spots are fatal to the main contribution. Since rotation matrices preserve Euclidean norms, U_r = (k_r/2)||p_i-o_k||^2 exactly; the angle alpha cancels and no rotational steering force exists. Eq. (24) is not the gradient of that potential—it is a rotated radial vector field, which has nonzero curl and cannot be a gradient. The chain rule also matters because alpha depends on position. The same error appears in Eq. (12), where the gradient of the planar weighted-norm potential is misstated; it should include k_v^T k_v(p_i-o_k)/||k_v(p_i-o_k)||. So even the planar obstacle avoidance law is not the gradient of the stated potential.\n\nThe simulations are single-run, use hand-tuned gains, have no baselines, and no code or data are provided. They show behavior, but they don't validate the specific claim of 3D maneuverability. The formation-control experiments might be fine on their own, but they reuse known methods.\n\nBottom line: this paper should not be published as is. The central novelty is vacuous, and the gradient errors are elementary. A reject is appropriate. If the authors fix the math and reframe the contribution honestly, there might be a small incremental paper on Lloyd-based formation control with potential-field avoidance. For now, I wouldn't cite it or spend referee time on it.","headline":"The 3D avoidance term is rotation-invariant, so the central claim of a 3D maneuver collapses; the formation control part is standard and the simulations don't compensate.","tokens_in":669,"tokens_out":985,"would_cite":false,"duration_ms":38684,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a semi-centralized controller that combines a probabilistic Lloyd's algorithm for formation with distributed potential fields for collision and obstacle avoidance, extending pigeon-inspired maneuvers to 3D for…","keywords":["multi-UAV systems","swarm intelligence","formation control","collision avoidance","obstacle avoidance","centroidal Voronoi tessellation","Lloyd's algorithm","distributed control"],"falsifier":"Evaluate the rotational potential in Eq. (20) for any detected obstacle: because $T_{rx},T_{ry},T_{rz}$ are rotations, $\\|T_{rx}T_{ry}T_{rz}(\\boldsymbol{p}_i-\\boldsymbol{o}_k)\\|=\\|\\boldsymbol{p}_i-\\boldsymbol{o}_k\\|$, so $U_r=\\frac{k_r}{2}\\|\\boldsymbol{p}_i-\\boldsymbol{o}_k\\|^2$ and its gradient is the radial force $k_r(\\boldsymbol{p}_i-\\boldsymbol{o}_k)$, independent of the angle $\\alpha$. A simulation with only the obstacle term activated should therefore show no lateral or vertical steering around an off-axis obstacle, which would contradict the reported 3D obstacle avoidance maneuver.","tokens_in":9665,"feed_emoji":"🚁","tokens_out":8925,"duration_ms":77769,"temperature":0.7,"pith_summary":"This paper aims to show that a multi-UAV swarm can keep a formation, avoid hitting each other, and steer around static and moving obstacles in 3D using a controller that mixes two inspirations: territorial tilapia fish for formation geometry and pigeons for obstacle maneuvers. The formation layer is semi-centralized, using a probabilistic Lloyd's algorithm to compute centroidal Voronoi positions that the drones track. The safety layer is distributed, with each drone applying local potential-field controls for inter-vehicle collision avoidance and obstacle avoidance. The proposed 3D obstacle term rotates the obstacle-relative displacement with yaw and pitch rotation matrices inside a potential function, which is meant to give drones extra maneuverability beyond planar avoidance. The paper reports simulations of 8 and 12 drones, including formation change and recovery between buildings and around static and dynamic obstacles, and claims these runs stay collision-free without consensus-based control.","feed_headline":"Fish-and-pigeon controller steers 12 drones through 3D obstacles","feed_subtitle":"A central planner sets formation targets while each drone avoids collisions locally, dropping consensus complexity.","key_machinery":"The carrying mechanisms are the probabilistic Lloyd's algorithm for centroidal Voronoi tessellation, which produces the formation targets, and the rotational obstacle potential $U_r=\\frac{k_r}{2}\\|T_{rx}T_{ry}T_{rz}(\\boldsymbol{p}_i-\\boldsymbol{o}_k)\\|^2$ with yaw and pitch rotation matrices $T_{ry}$, $T_{rz}$ and an identity matrix $T_{rx}$; the distance-dependent angle $\\alpha$ is supposed to shape the avoidance force as the drone approaches an obstacle. These are combined through the superposition control law $\\boldsymbol{u}_i=\\boldsymbol{u}_{fi}+\\boldsymbol{u}_{ci}+\\boldsymbol{u}_{oi}$, with the obstacle term gated by two detection conditions: range within $r_d+r_{ok}$ and the obstacle lying inside a conical field of view around the flight direction.","core_discovery":"The central claim is that formation planning and safety can be cleanly separated in a multi-UAV swarm: Lloyd's algorithm on a Voronoi cost function supplies each drone with an optimal target position, while a superposition of three control terms—formation tracking, inter-vehicle repulsion, and an obstacle potential—keeps the flight collision-free. The distinct novelty is the obstacle term, written as a rotational potential $U_r=\\frac{k_r}{2}\\|T_{rx}T_{ry}T_{rz}(\\boldsymbol{p}_i-\\boldsymbol{o}_k)\\|^2$, whose gradient is intended to push drones around obstacles in 3D rather than simply away from them; the rotation angle $\\alpha$ decreases from $\\pi/2$ at the safety radius to zero at the detection radius. Applied to 8 and then 12 UAVs, the controller is presented as achieving safe formation change, passage through building gaps, avoidance of static and dynamic obstacles, and formation recovery, with inter-drone distances held above the safety threshold.","pith_inferences":["Editorially, the 3D claim stands or falls on whether the rotation matrices change the force direction; because rotation preserves vector norms, the potential as written is radial, so a genuine 3D maneuver needs an additional mechanism such as heading-dependent asymmetry in the potential.","A testable extension is to replace the rotational potential with an explicit tangential steering term and compare minimum-distance statistics against the present formulation in the same obstacle layouts.","The modular architecture suggests a practical deployment principle: formation targets can update at a slower global rate while avoidance runs at a fast local rate; a hardware or high-fidelity simulator test could measure how much update-rate separation the safety distances tolerate."],"forward_implications":["Multi-UAV swarms could perform formation changes and recovery in cluttered urban areas without consensus-based coordination, lowering controller design complexity.","The centralized/distributed split scales to at least 12 UAVs in the presented simulations, with only local neighbor and obstacle information used by the safety layer.","Dynamic obstacles can be handled online by range and field-of-view detection, rather than by precomputed global path planning.","The same controller structure could be reused for non-static barrier areas, since the Lloyd targets are recomputed as the barrier moves."],"supporting_citations":[{"why":"Supplies the collision-avoidance law and the planar pigeon-inspired obstacle controller that the paper extends to 3D.","marker":"[3]"},{"why":"Provides the biological model of pigeons' visually guided obstacle flight that motivates the detection and field-of-view design.","marker":"[15]"},{"why":"Demonstrates pigeon-inspired avoidance on four quadrotors, the baseline the paper scales up to 12 UAVs.","marker":"[16]"},{"why":"Provides the probabilistic Lloyd's algorithm used to compute centroidal Voronoi formations.","marker":"[19]"},{"why":"Supports the Voronoi/coverage approach for 3D UAV navigation and flocking.","marker":"[20]"},{"why":"Grounds the multicenter cost function and centroidal Voronoi configuration for optimal deployment.","marker":"[21]"},{"why":"Provides the rotation-matrix formulation used for 3D attitude and avoidance maneuvers.","marker":"[25]"}],"fun_headline_variants":["Fish and pigeon brains guide drone swarms around 3D obstacles","Nature-inspired swarm flies 12 drones safely through 3D obstacles","Swarm uses fish and pigeon instincts for 3D obstacle dodging","Central planner sets formation and local repulsion avoids 3D collisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that rotating the obstacle-relative displacement inside the avoidance potential changes the potential's value and force direction enough to add 3D maneuverability; if the rotation leaves the norm unchanged, the avoidance term collapses to radial repulsion and the distinctive 3D claim loses its mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Fish and pigeon brains guide drone swarms around 3D obstacles","Nature-inspired swarm flies 12 drones safely through 3D obstacles","Swarm uses fish and pigeon instincts for 3D obstacle dodging","Central planner sets formation and local repulsion avoids 3D collisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001026,"raw_usage":{"total_tokens":4310,"prompt_tokens":914,"completion_tokens":3396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":3320}},"tokens_in":530,"tokens_out":3396,"duration_ms":24860,"temperature":1.0,"reasoning_tokens":3320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:05:05.621535+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the rotational potential in Eq. (20) for any detected obstacle: because $T_{rx},T_{ry},T_{rz}$ are rotations, $\\|T_{rx}T_{ry}T_{rz}(\\boldsymbol{p}_i-\\boldsymbol{o}_k)\\|=\\|\\boldsymbol{p}_i-\\boldsymbol{o}_k\\|$, so $U_r=\\frac{k_r}{2}\\|\\boldsymbol{p}_i-\\boldsymbol{o}_k\\|^2$ and its gradient is the radial force $k_r(\\boldsymbol{p}_i-\\boldsymbol{o}_k)$, independent of the angle $\\alpha$. A simulation with only the obstacle term activated should therefore show no lateral or vertical steering around an off-axis obstacle, which would contradict the reported 3D obstacle avoidance maneuver.","supporting_citations":[{"cited_title":"Live -fly experimentation for pigeon-inspired obstacle avoidance of quadrotor unmanned aerial vehicles,","cited_arxiv_id":null,"evidence_quote":"Demonstrates pigeon-inspired avoidance on four quadrotors, the baseline the paper scales up to 12 UAVs."},{"cited_title":"Computationally-efficient distributed algorithms of navigation of teams of autonomous UAVs for 3D coverage and flocking,","cited_arxiv_id":null,"evidence_quote":"Supports the Voronoi/coverage approach for 3D UAV navigation and flocking."}],"review_version":1}