{"id":"61883f06-1deb-49ec-ba16-f7bcd6266459","arxiv_id":"2508.21316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A UAV formation follows a path, senses obstacles via ISAC signals, repositions on its virtual circle to lower the CRLB, and fuses avoidance with path-following through null-space priorities.","lead":"The paper proposes a combined scheme for UAV formations to follow a virtual path while sensing and avoiding dynamic obstacles, using 5G-style ISAC signals, a variable formation that repositions to improve sensing accuracy, and a priority hierarchy that keeps avoidance safe. If the simulations hold, it offers a modular way to coordinate these tasks without retraining the path-following policy after each avoidance event.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Conflict-free scheduling of sensing and avoidance is asserted but never simulated concurrently; fusion law (61) simply adds two equal-priority commands.","rationale":"The reader's weakest assumption centers on the CRLB and the linear-Gauss-Markov obstacle model, which is a legitimate structural risk for the sensing gain claim. My concern is adjacent but distinct: even if the CRLB machinery is correct, the paper's headline scheduling claim—'conflict-free scheduling of the three subtasks'—is only validated under non-overlapping subtask activation. The NSB architecture in Eq. (61) treats sensing and avoidance as equal-priority and simply adds their velocities, with no null-space projection between them and no reported weights k1,k2. In the simulation, VFEO runs at 276/278 s and avoidance at 281–330 s, so the potentially conflicting overlap is never tested. This does not show the method is wrong; it shows the strongest claim is under-supported. The reader's verdict of CONDITIONAL already captures the need for more validation, so I do not change the verdict. A targeted simulation with simultaneous sensing reconfiguration and avoidance would either close the gap or expose a real conflict.","tokens_in":23184,"tokens_out":3898,"duration_ms":43442,"concrete_test":"Extend the Section VI-C scenario so the obstacle enters rs during the VFEO window (e.g., shift s1 so the predicted distance at 276 s is just below rs), run the full N-HSF law with the paper's stated parameters, and report min_i ||u_i(k)-s(k)|| for 270–290 s and whether the VFEO solution satisfies (57e). Then sweep (k1,k2) over [0,1] with k1+k2=1; if any combination yields min distance < rs or degrades the CRLB improvement, the 'conflict-free' conclusion requires revision. Also report the k1,k2 values used in Fig. 18.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VII concludes that N-HSF 'ensures no conflict of subtasks,' but the evidence in Section VI-C only exercises sensing+path-following (VFEO at 276/278 s) and avoidance+path-following (281–330 s) separately; the three subtasks are never active simultaneously. The fusion law (Eq. 61) gives obstacle avoidance and sensing equal highest priority and does not project either into the other's null space: v = k1*v1 + k2*v2 + (I - J1†J1)*v3. No values for k1,k2 are reported, and there is no arbitration rule when v1 and v2 point in conflicting directions. Although the VFEO optimization includes constraint (57e) (distance to the predicted obstacle >= rs at k+1), the final control is the weighted sum, not the optimizer output alone; the sum can in principle cancel or reduce the safety margin of v1. Because the simultaneous case is absent from the simulation, the paper's strongest scheduling claim is unsupported. This is a load-bearing gap rather than a mere lack of ablations: the proposed architecture has no hierarchy between two 'highest priority' tasks.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-UAV formation path-following, cooperative sensing, and obstacle avoidance architecture for 3D space. Three subtasks are addressed: (i) a DDPG-based path-following policy with an adaptive reward that weights distance and velocity errors (AWPF); (ii) an ISAC-signal-based obstacle position/velocity estimator whose Cramér-Rao lower bound (CRLB) drives a variable formation optimization (VFEO) that repositions UAVs on a virtual circle to reduce the predicted position error; and (iii) an online, pre-training-free obstacle avoidance law. The subtask commands are fused by a null-space-based hierarchical strategy (N-HSF) that assigns top priority to obstacle avoidance and obstacle sensing and projects path-following into the null space of avoidance. Simulations with five UAVs and two dynamic obstacles demonstrate path-following accuracy, improved obstacle position estimates after two formation reconfigurations, and collision-free avoidance, and the paper concludes that the fusion strategy realizes conflict-free scheduling of the three subtasks.","tokens_in":23496,"tokens_out":4920,"duration_ms":49252,"significance":"If fully substantiated, the paper would offer a useful integration of DRL-based path-following, CRLB-optimized cooperative sensing, and null-space control for UAV formations. The AWPF design is a sensible extension of existing reward shaping, and the per-UAV trained policy with virtual-leader-based formation control is a practical scalability feature. The CRLB derivation for non-uniform DM-RS signals and the VFEO optimization are technically concrete, and the complexity analyses for AWPF and VFEO are a valuable addition. The strongest advertised claim, however, is the conflict-free fusion of all three subtasks. That claim is currently not supported by the evidence because the three subtasks are never exercised simultaneously, and the fusion law lacks an arbitration rule between the two equal-priority commands. The paper's single-scenario simulations also leave the quantitative improvements vulnerable to tuning artifacts. With additional validation, the architecture could be a solid contribution; in its present form, the central scheduling claim remains unproven.","major_comments":[{"comment":"The central claim that N-HSF 'possesses the ability to realize conflict-free scheduling of the three subtasks' (conclusion, Section VII; also asserted in Section VI-C) is not supported by the evidence. In the simulation, sensing+path-following (276/278 s) and avoidance+path-following (281–330 s) are tested separately; the three subtasks are never active simultaneously. The fusion law (61) gives v1 (avoidance) and v2 (sensing) equal highest priority and projects only v3 into the null space of v1. No projection, null-space ordering, or arbitration is defined between v1 and v2, and k1, k2 are not reported. Because the final command is a weighted sum, v1 and v2 can in principle point in conflicting directions and partially cancel; the safety margin implied by constraint (57e) in the VFEO optimization is not preserved in the fused command. A simulation with all three subtasks active, includin","section":"§V, Eq. (61); §VI-C"},{"comment":"All quantitative conclusions rest on a single simulated scenario: five UAVs, two obstacles, one trajectory per phase. The reported 21–124% AWPF accuracy improvement and the reductions of obstacle position error from 2.05 m to 0.02 m and from 0.78 m to 0.03 m at 276 s and 278 s are single-run values. No Monte Carlo trials, confidence intervals, or sensitivity analyses are provided with respect to the hand-tuned quantities that drive the algorithm: the distance threshold 3 m and constant 40 in (62)–(63), and the positioning accuracy threshold ζP_CRLB = 0.5 m in Table I. Since the AWPF reward schedule and the VFEO trigger are both governed by these parameters, statistical or sensitivity evidence is necessary to establish that the improvements are not artifacts of a particular tuning.","section":"§VI, Figs. 10–18"},{"comment":"The CRLB used in the VFEO objective replaces the true ranges r_o^i with the arithmetic mean of multiple measurements, and the predicted obstacle position at k+1 is obtained from the linear Gauss-Markov model (17) with Gaussian driving noise. If the obstacle motion deviates from (17), or if the strict clock synchronization and Gaussian noise assumptions stated in Section III-C fail, the computed CRLB is not the actual error lower bound. Consequently, the optimized formation from (57) can be suboptimal and the trigger εP(k+1) > ζP_CRLB can be unreliable. This assumption is structural because it underlies both the sensing objective and the reconfiguration condition. A robustness test with model mismatch and imperfect synchronization would clarify the domain of validity of the claimed decimeter-level sensing improvement.","section":"§III-D, Eqs. (46)–(47)"}],"minor_comments":[{"comment":"The text states ω2 ∝ 1/ed,ie, but the explicit schedule in (62)–(63) is not a direct proportionality: ω2 = 40/(ed,ie+40) for ed,ie > 3 and 0.95 otherwise. Please reconcile the description with the formula, and use a proper summation notation in ∑ ωo = 1.","section":"§IV-A, Eq. (52)"},{"comment":"The peak indices L_{s_i,w} and L_{s_i,q} are used before being explicitly defined. Define them or state clearly that they follow the peak-search procedure of [24].","section":"§III-A, Eqs. (23)–(24)"},{"comment":"The legend entry 'VAV3 VF' appears to be a typo for 'UAV3 VF'.","section":"Fig. 11"},{"comment":"The statement that sensing is equivalently completed by changing VFTs is not fully consistent with v2 being a control velocity in the fusion law (61). Clarify whether the sensing subtask contributes a velocity command or only a VFT update.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper's main claimed contribution—conflict-free hierarchical fusion—is also its weakest supported point. I would ask for a simultaneous-subtask simulation and for an explicit arbitration rule or null-space ordering between obstacle avoidance and sensing. The narrow single-scenario validation should be expanded, at least with a few Monte Carlo runs or sensitivity sweeps over the key thresholds. The ISAC/CRLB portion is competently developed and could be publishable after these additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the variable-formation idea—repositioning UAVs on the virtual circle to minimize the sensing CRLB while keeping the formation—is the real contribution. The rest is mostly known machinery (DDPG reward shaping, TWLS for TDOA/FDOA, NSB fusion), but assembled carefully and with honest complexity analysis. The paper is not circular: CRLB and TWLS are benchmarked against external theory, and the adaptive-weight reward is a modest heuristic, not an invented law.\n\nWhat I found well done: the CRLB derivation from non-uniform DM-RS is explicit, the VFEO constraint set (formation radius, speed, separation, obstacle distance) is sensible, and the simulation shows striking sensing improvements at the two reconfiguration instants (2.05 m to 0.02 m, 0.78 m to 0.03 m). The DRL path-following claims a 21–124% improvement over fixed-weight baselines; that's plausible given the baselines are deliberately weaker.\n\nThe soft spots are in the validation, not the machinery. One scenario, five UAVs, two obstacle encounters, no Monte Carlo, no error bars. Thresholds (3 m, 40, ζ_CRLB=0.5) are hand-picked, and k1,k2 in the fusion law are never reported. More importantly, the paper's concluding claim that N-HSF 'ensures no conflict of subtasks' is not backed by the experiments: sensing+path-following and avoidance+path-following are exercised separately; the three subtasks never run simultaneously. And equation (61) treats sensing and avoidance as equal-priority and just adds their commands—no null-space projection between them, no arbitration when they point opposite directions. So the load-bearing scheduling claim is unsupported as it stands. That's a genuine gap, not a nitpick.\n\nAlso, the CRLB in (42)–(49) uses the arithmetic mean of measurements in place of true ranges; under model mismatch or clock drift the optimized geometry can be off. But that's a standard practical caveat and not fatal to the integration.\n\nWho benefits: this is useful for people working on ISAC-equipped UAV formations and DRL formation control. It deserves a serious referee: the idea is new, the derivation is careful, and the gap is fixable by running the simultaneous case, reporting fusion weights, and adding sensitivity analysis. I'd send it to review, expecting major revision.","headline":"Solid integration with one genuinely new variable-formation sensing idea, but the 'conflict-free three-subtask scheduling' claim is not actually simulated.","tokens_in":23975,"tokens_out":2267,"would_cite":true,"duration_ms":24130,"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 UAV formation can follow a prescribed path, sense a moving obstacle to centimeter-level accuracy by reconfiguring its shape, and avoid collisions, with the three tasks scheduled conflict-free through a null-space pr","keywords":["UAV formation","path-following","integrated sensing and communication","Cramér-Rao lower bound","variable formation","obstacle avoidance","null-space-based fusion","deep reinforcement learning"],"falsifier":"Run the VFEO optimization in simulation or field tests with two obstacle trajectories: one matching the Gauss-Markov motion model (17) and one with an abrupt turn or acceleration burst. Measure the actual obstacle-position RMSE around the reconfiguration instants. If the maneuvering trajectory fails to reproduce the 0.02–0.03 m errors or breaks the collision-free guarantee, while the model-matched trajectory succeeds, the smooth-motion/Gaussian premise is the load-bearing assumption.","tokens_in":1622,"feed_emoji":"📡","tokens_out":2649,"duration_ms":93039,"temperature":0.7,"pith_summary":"The paper tries to prove that three competing jobs in a UAV formation mission—staying on a path, estimating where a moving obstacle is, and steering clear of it—can be handled by one coordinated system instead of three separately retrained models. Its central idea is to reuse the same ISAC signals for communication and sensing, and to let the formation change shape exactly when the predicted Cramér-Rao lower bound on obstacle position exceeds a threshold. The variable-formation algorithm chooses new UAV positions on the virtual circle that minimize that bound, while a null-space-based hierarchy gives sensing and avoidance the highest priority and lets path-following continue in their null space. Simulations show obstacle-position error dropping from 2.05 m to 0.02 m and from 0.78 m to 0.03 m at the two reconfiguration instants, along with collision-free avoidance in both chasing and following phases. If right, the result is a path-following controller that needs no retraining after avoidances and gets sharper obstacle estimates exactly when they matter.","feed_headline":"Formation reshape cuts obstacle error from 2 m to 2 cm","feed_subtitle":"Null-space fusion lets a UAV team follow a path, dodge obstacles, and sharpen sensing without retraining.","key_machinery":"CRLB-based reconfiguration and null-space fusion. The Cramér-Rao lower bound ε_P = sqrt(tr(CRLB_P)) from equations (42)-(51) converts formation geometry into a predicted sensing error, and the VFEO optimization (57) minimizes it over UAV positions on the virtual circle. The null-space-based fusion rule (61), v = k1 v1 + k2 v2 + (I − J†_1 J_1)v3, assigns obstacle avoidance and sensing the highest priority and projects the DRL path-following velocity into their null space, so path-following continues partially even while obstacles are being handled.","core_discovery":"The paper's central claim is that formation geometry itself is a sensing resource. Starting from range and radial-velocity measurements obtained from ISAC signals, the authors derive the Cramér-Rao lower bound for the fused obstacle-position estimate and show that this bound depends on where the UAVs sit on the formation circle. The VFEO algorithm turns that bound into an objective: at time k, if the predicted next-step position error ε_P(k+1) exceeds the threshold ζ_CRLB, it solves a constrained optimization for new UAV positions, subject to staying on the virtual circle, keeping altitude, respecting speed limits, and maintaining safety distances. In simulation the reconfiguration reduces t","pith_inferences":["The same CRLB-minimizing reconfiguration idea could be applied to other cooperative localization tasks, such as tracking multiple maneuvering targets or ground users, wherever formation shape can be adjusted; the paper itself only treats a single obstacle.","The fixed threshold ζ_CRLB = 0.5 m could be made adaptive to mission phase or threat level, since the paper does not explore how the threshold choice trades sensing accuracy against formation stability.","A testable extension is to compare VFEO's optimized placements against random or equidistant formation placements under identical measurement noise, isolating how much of the error drop comes from the optimized geometry rather than from simply taking fresh measurements.","The nested null-space structure could be extended to additional subtasks, such as maintaining communication links or energy constraints, by stacking further projections; the paper stops at three subtasks."],"forward_implications":["A formation can improve obstacle localization without adding sensors or retraining, simply by redistributing UAV positions on the same formation circle.","The DRL policy, trained for a single UAV and driven by master-assigned virtual following targets, scales naturally to different numbers of UAVs.","After an obstacle-avoidance episode, the formation's path-following error returns to its pre-avoidance level, indicating that avoidance does not permanently disturb the learned following behavior.","Using ISAC signals for both communication and sensing reduces communication overhead compared with systems that separate the two functions.","The CRLB threshold provides an online trigger for reconfiguration: the formation reshapes only when predicted sensing error exceeds a set bound."],"supporting_citations":[{"why":"Supplies the 5G PRS/DM-RS sensing framework and the CRLB derivation for range and radial-velocity estimates used in Section III.","marker":"[24]"},{"why":"Provides the two-step weighted least squares TDOA/FDOA solution used to fuse multi-UAV range and radial-velocity measurements into obstacle position and velocity.","marker":"[34]"},{"why":"Gives the OFDM positioning CRLB approach that the paper extends to formation-level obstacle sensing.","marker":"[35]"},{"why":"Provides the joint TDOA/FDOA estimation bound that motivates fusing both range and radial velocity for obstacle localization.","marker":"[33]"},{"why":"Supplies the null-space-based trajectory tracking and collision avoidance method underlying the hierarchical fusion rule (61).","marker":"[38]"},{"why":"Baseline DDPG formation path-following method whose reward design the AWPF algorithm extends with adaptive weights.","marker":"[7]"},{"why":"Provides the discrete-time obstacle dynamics model and CRLB-based coordinated control used for the sensing formulation.","marker":"[32]"},{"why":"Provides the OFDM signal model and joint sensing-communication cooperative sensing framework used for the ISAC waveform.","marker":"[30]"}],"fun_headline_variants":["Formation geometry becomes a sensor: obstacle error drops 100x","UAVs reshape on the fly to cut obstacle localization error 100x","Sensing gain from formation shape: 2 m error to 2 cm without retraining","Variable formation turns geometry into a sensing boost for UAVs","No retraining needed: UAVs reshape formation to sense and dodge"],"cache_read_input_tokens":25728,"weakest_assumption_plain":"The entire sensing gain rests on the obstacle's motion being smooth enough to be predicted by a linear Gaussian model, on strict clock synchronization between UAVs, and on treating averaged measurements as the true ranges in the error bound—if any of these fails, the reshaped formation may not actually deliver the reported accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Formation geometry becomes a sensor: obstacle error drops 100x","UAVs reshape on the fly to cut obstacle localization error 100x","Sensing gain from formation shape: 2 m error to 2 cm without retraining","Variable formation turns geometry into a sensing boost for UAVs","No retraining needed: UAVs reshape formation to sense and dodge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000484,"raw_usage":{"total_tokens":2226,"prompt_tokens":745,"completion_tokens":1481,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":1384}},"tokens_in":489,"tokens_out":1481,"duration_ms":12242,"temperature":1.0,"reasoning_tokens":1384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:23:28.730178+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the VFEO optimization in simulation or field tests with two obstacle trajectories: one matching the Gauss-Markov motion model (17) and one with an abrupt turn or acceleration burst. Measure the actual obstacle-position RMSE around the reconfiguration instants. If the maneuvering trajectory fails to reproduce the 0.02–0.03 m errors or breaks the collision-free guarantee, while the model-matched trajectory succeeds, the smooth-motion/Gaussian premise is the load-bearing assumption.","supporting_citations":[{"cited_title":"5G PRS-based sensing: A sensing reference signal approach for joint sensing and communication system,","cited_arxiv_id":null,"evidence_quote":"Supplies the 5G PRS/DM-RS sensing framework and the CRLB derivation for range and radial-velocity estimates used in Section III."},{"cited_title":"An accurate algebraic solution for moving source location using TDOA and FDOA measurements,","cited_arxiv_id":null,"evidence_quote":"Provides the two-step weighted least squares TDOA/FDOA solution used to fuse multi-UAV range and radial-velocity measurements into obstacle position and velocity."},{"cited_title":"Achievable localization accuracy of the posi- tioning reference signal of 3GPP LTE,","cited_arxiv_id":null,"evidence_quote":"Gives the OFDM positioning CRLB approach that the paper extends to formation-level obstacle sensing."},{"cited_title":"Joint TDOA and FDOA estimation: A conditional bound and its use for optimally weighted localization,","cited_arxiv_id":null,"evidence_quote":"Provides the joint TDOA/FDOA estimation bound that motivates fusing both range and radial velocity for obstacle localization."},{"cited_title":"A novel null-space-based UA V trajectory tracking controller with collision avoidance,","cited_arxiv_id":null,"evidence_quote":"Supplies the null-space-based trajectory tracking and collision avoidance method underlying the hierarchical fusion rule (61)."},{"cited_title":"USV formation and path-following control via deep reinforcement learning with random braking,","cited_arxiv_id":null,"evidence_quote":"Baseline DDPG formation path-following method whose reward design the AWPF algorithm extends with adaptive weights."},{"cited_title":"Coordinated CRLB-based control for tracking multiple first responders in 3D en- vironments,","cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time obstacle dynamics model and CRLB-based coordinated control used for the sensing formulation."},{"cited_title":"Performance of joint sensing-communication cooperative sensing UA V network,","cited_arxiv_id":null,"evidence_quote":"Provides the OFDM signal model and joint sensing-communication cooperative sensing framework used for the ISAC waveform."}],"review_version":1}