{"id":"c46ed459-9a59-482a-9c9a-d77f826f0235","arxiv_id":"2505.07523","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Parallel tuning of MAV controller gains via averaging or gain distribution is demonstrated, with reduced variance in simulation and two successful outdoor flights.","lead":"A team of robotics researchers shows how to tune drone altitude controller gains faster by running the tuning procedure on several drones at once, either averaging their measurements or testing different gain settings in parallel. The approach is model-free and deterministic in runtime, and was tested in simulation and on two real drones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-AVG time reduction relies on comparing costs from different MAVs under an untested identical-MAV assumption; with a very flat cost surface, small systematic platform or wind differences can flip EQL decisions.","rationale":"The reader's weakest assumption—that the N MAVs are effectively identical and differ only by measurement noise—is indeed the load-bearing condition for the parallel tuning argument. I agree with that identification and with the CONDITIONAL verdict. The most concrete way the assumption can break is in the noAVG mode, where the EQL update directly compares cost values from two physically distinct MAVs. Because the cost surface is explicitly described as extremely flat, even a small systematic offset between platforms or locations can change the sign of the comparison and send the EQL search to the wrong interval. This is not merely a scalability or statistical-power complaint: it threatens the validity of the time-reduction claim itself, since AVG does not reduce tuning time and noAVG is the only mode that does. The simulation experiments reuse identical Gazebo models, so they cannot expose this bias; the two real-world runs, one per mode, have no repeats and no sequential baseline, so they cannot confirm the claim either. The proposed test—adding a controlled bias in simulation and sweeping its magnitude—directly probes whether the method tolerates realistic between-MAV differences. If it does not, the central claim should be narrowed to controlled, well-calibrated fleets or to scenarios where MAVs are co-located. The paper remains a plausible engineering contribution, but the evidence currently supports only a conditional acceptance pending this check.","tokens_in":9307,"tokens_out":5838,"duration_ms":62637,"concrete_test":"In the Gazebo setup of Section IV, induce a controlled between-MAV bias by adding a constant delta to the slave MAV's altitude error (or by adding a 5% mass offset) and rerun the 10-repetition noAVG tuning for delta values from 0 up to the observed inter-run kD std (0.12). If the mean final gains shift by more than the AVG gain improvement (e.g., kP 0.8656 -> 0.9000) or the variance increases significantly, the cross-platform comparison is not robust and the real-world single runs do not validate the time-reduction claim. Optionally record single-MAV sequential EQL wall-clock as a baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract) is that the parallel procedure improves tuning while decreasing tuning time. The time decrease comes only from the 'without AVG' mode (Section III), which compares cost values J_j measured on different MAVs at different gain configurations and uses the sign of the difference to update the EQL interval. This comparison is valid only under the Section III assumption of 'N identical MAVs' and statistically identical environments; otherwise the cost difference between the two configurations is confounded with a between-platform offset. The paper's own data show why this matters: the cost surface in Fig. 2 is 'extremely flat' (log-scale, range roughly 1.4 over the normalized gain grid), while the 10-run Gazebo without-AVG tuning has kP std 0.0508 and kD std 0.1225. A systematic difference between MAVs—mass, sensor calibration, wind gradient—of the same order as the measurement noise can therefore flip the J(p(k-)) < J(p(k+)) comparisons that drive the whole EQL update. In the real-world runs (Figs. 5-7) the MAVs were 'placed farther away from one another', so spatial wind differences are likely; only one run per mode is reported, so the bias cannot be estimated. AVG mode is not a remedy for the time claim because the authors state it 'offers no reduction in tuning time'. Thus the 'decreasing tuning time' half of the central claim is supported only by the noAVG experiments, which are exactly the ones most exposed to this bias, and no baseline sequential tuning time is measured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a parallel tuning method for the altitude controller gains (kP, kD) of a swarm of MAVs, extending the EQL zero-order optimization algorithm. Two distribution strategies are presented: AVG, in which all MAVs test identical gains and the cost readings are averaged to reduce measurement noise, and noAVG, in which different gain configurations are tested on different MAVs simultaneously to reduce tuning time. The method is evaluated in Gazebo with 10 repetitions per mode, including a payload variant, and in two real-world outdoor flights, one per mode. The paper claims a reliable, time-deterministic, model-free tuning procedure that improves tuning results while decreasing tuning time.","tokens_in":9619,"tokens_out":4140,"duration_ms":39375,"significance":"If the claims were fully supported, the paper would make a useful practical contribution to UAV controller tuning: a deterministic-time, model-free procedure with open-source code and both simulated and real-world demonstrations. The AVG strategy shows a genuine, if modest, repeatability improvement in the Gazebo tests (kP standard deviation 0.0192 versus 0.0508 in the noAVG mode), which is a measurable positive result. The noAVG idea is also appealing in principle. However, the evidence for the headline 'decreasing tuning time' claim and for the reliability of the noAVG mode is currently incomplete, as detailed below.","major_comments":[{"comment":"The noAVG mode compares cost-function values J_j measured on different MAVs at different gain configurations to decide the EQL interval update (cases 2a and 2b in Section III). This is valid only under the stated assumption of N identical MAVs in statistically identical environments. The paper does not validate this assumption: Fig. 2 shows a very flat cost surface, the Gazebo noAVG results have non-negligible dispersion (kP = 0.8656 ± 0.0508, kD = 0.3094 ± 0.1225), and the real-world runs place the MAVs far apart (Fig. 5) with one run per mode. A small systematic platform or wind bias can therefore flip the J comparisons that drive the update, so the time-reduction mode is exactly the mode most exposed to bias. This issue is load-bearing because the AVG mode is explicitly said to offer no reduction in tuning time.","section":"Section III (EQL update rule, noAVG mode)"},{"comment":"The central claim of decreased tuning time is not quantified against any sequential or single-MAV baseline. The paper reports a 560 s experiment for the parallel procedure, but no wall-clock or per-iteration comparison with sequential EQL tuning under comparable conditions is provided. The near-40% figure in Section III refers to reusing a previously evaluated point within one EQL bootstrap, not to a measured speedup of the parallel swarm procedure. Without a baseline, the abstract's 'decreasing the tuning time' claim is not supported by the presented data.","section":"Section IV (experimental results)"},{"comment":"The evaluation of 'improving the results' uses the same altitude-tracking cost function that the tuning procedure optimizes, and no independent validation is provided on a different trajectory, metric, or set of conditions. This matters because the cost surface in Fig. 2 is described as 'extremely flat', so gains differing by the reported dispersions may be statistically indistinguishable in terms of the tuning objective. The paper would need either a comparison against the known grid optimum in Fig. 2, a different validation metric, or a statistical significance test to substantiate the improvement claim.","section":"Section IV (Figs. 2, 6, 7)"},{"comment":"The real-world evidence consists of two single runs, one per mode, with no repetitions, error bars, or significance tests. Given that the paper claims a 'reliable tuning mechanism', the real-world portion alone cannot support that claim. This is particularly relevant because the two modes are not directly comparable in wall-clock time: Fig. 6 spans 0–600 s while Fig. 7 begins at 250 s, and no total duration or stopping criterion is stated for either run.","section":"Section IV (real-world evaluation)"}],"minor_comments":[{"comment":"The caption reads '2 MAV models in the simulation (left)' in both clause positions; one instance should likely refer to the right panel.","section":"Fig. 3 caption"},{"comment":"The step 'put k_D^{(k+1)} = k_D^{(k)}' appears self-referential; it presumably intends to fix one parameter while optimizing the other, but the notation should be clarified.","section":"Section III (bootstrap step)"},{"comment":"The state-machine description is given for N=2 only, although the method is claimed to scale to arbitrary even N; a brief description of how the synchronization and optimization steps generalize would improve clarity.","section":"Section III (state machine)"},{"comment":"There are numerous typographical and formatting issues, including the dangling '+' at the end of Eq. (2), irregular equation numbering, and awkward phrasing in the introduction; a careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable after revision if the authors add a sequential baseline, validate the noAVG mode with repeated runs and ideally an identical-MAV check, and provide independent validation of the tuned gains. The flat cost surface also raises questions about whether the tuning problem itself is well-posed for demonstrating improvement, so an explicit comparison with the grid optimum or a different metric would strengthen the claims considerably."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, modest engineering extension of the group's earlier EQL tuning work to parallel MAV swarms. The AVG strategy measurably reduces variance in simulation; the time-reduction claim for the noAVG strategy is plausible but not actually demonstrated.\n\nWhat's new: two concrete distribution strategies (averaging identical-gain costs across agents, and splitting gain configurations across agents), with the EQL iteration count as a deterministic time model. The simulation campaign with 10 repetitions is a real plus, and the code being public helps reproducibility. The variance reduction from AVG in the Gazebo tests is clear (kP std drops from 0.051 to 0.019), so the reliability claim has a leg to stand on.\n\nThe soft spots are mostly about evidence weight. The 'decreasing tuning time' part of the abstract leans entirely on the noAVG mode, but no wall-clock baseline against sequential EQL is reported; the speedup is inferred from the algorithm's iteration count. More importantly, the noAVG cost comparison uses J values from different MAVs at different gain configurations, and the paper's own Figure 2 shows a very flat cost surface. The stress-test note is right: a small systematic bias between platforms—mass, sensor calibration, wind gradient—can flip the sign of the cost difference and send the EQL update the wrong way. The identical-MAV assumption is stated but never validated, and the real-world experiments are one run per mode with no error bars. So the central claim of 'improved results while decreasing tuning time' is only partially supported by the data as presented.\n\nAlso, scalability beyond N=2 is asserted but not shown, and the paper doesn't compare against any alternative parallel tuning baseline.\n\nWho this is for: researchers working on UAV controller auto-tuning, especially those doing field deployment. It deserves a serious referee—the method is cheap, deterministic, and the simulation evidence is honest—but the authors should be pushed to add a sequential baseline, more real-world repetitions, and at least a sanity check on the identical-MAV assumption.","headline":"A useful but under-validated extension of EQL tuning to swarms: the averaging mode is supported, the time-saving mode is plausible but lacks a measured baseline and is exposed to between-platform bias.","tokens_in":10168,"tokens_out":2380,"would_cite":false,"duration_ms":22134,"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 swarm of effectively identical MAVs can tune its altitude-controller gains in parallel, either suppressing noise by averaging cost readings or cutting tuning time by splitting the gain search.","keywords":["controller gain tuning","parallel tuning","swarm of MAVs","model-free optimization","equal-division method","zero-order optimization","altitude control","performance index averaging"],"falsifier":"Run the AVG mode repeatedly with two vehicles and then with four vehicles under the same outdoor conditions; if the standard deviation of the final tuned gains does not decrease as the number of vehicles grows, the claimed noise-reduction benefit is not real. Alternatively, fly the no-AVG mode with two vehicles of measurably different mass and check whether the gain that is best for the vehicle that tested it is also best for the other vehicle; if not, parallel testing is trading correctness for speed.","tokens_in":9150,"feed_emoji":"🚁","tokens_out":13284,"duration_ms":114821,"temperature":0.7,"pith_summary":"The paper is trying to establish that tuning the altitude-controller gains of a micro aerial vehicle need not be a slow, sequential, trial-and-error exercise: if several effectively identical vehicles are available, the search over gains can be distributed across them so that the swarm either reaches a good gain set sooner or reaches a more repeatable one under noisy measurements. The proposed procedure is model-free, uses only basic sensory data, and has a running time that is fixed in advance, because the underlying equal-division (EQL) search needs a known number of cost evaluations per parameter. Two distribution rules are compared: publishing the same gains to every vehicle and averaging the reported cost functions to suppress noise, and publishing different gains to different vehicles to test several configurations in parallel and shorten the tuning horizon. The claim is supported by simulation campaigns and by real outdoor flights with a pair of hexarotors, and it matters because controller tuning in practice can consume days, while a deterministic parallel procedure could be run in the field, even in stages between battery swaps.","feed_headline":"Parallel drones tune their controllers faster and with less noise","feed_subtitle":"Averaging gains across identical drones trims noise; splitting gain tests across them trims tuning time.","key_machinery":"The load-bearing mechanism is the equal-division (EQL) zero-order search, a bracketing rule that at each iteration evaluates two interior gain values inside the current uncertainty interval and narrows the interval according to which value gives the lower cost. It is run as a bootstrapped one-parameter-at-a-time loop alternating between the two altitude gains, and it carries over one of the two test points to the next iteration, so after the first round only a single new cost evaluation is needed per parameter, saving nearly 40% of the experiments. On the control side, the machinery is a three-state state machine on the master vehicle -- IDLE, FLYING, OPTIMIZE -- synchronized by semaphores, which parcels out the reference primitives, collects each vehicle's accumulated cost $J^{(i)}_j$, and publishes the updated gains; the distribution rule for those updated gains (identical-to-all with averaging, or different-per-vehicle) is what turns a single-vehicle tuner into a parallel swarm tuner.","core_discovery":"In the paper's own terms, the discovery is that the EQL zero-order tuning loop, originally designed for a single vehicle, can be executed by a swarm without changing the search logic: a master vehicle runs the optimization state machine, publishes gain updates to all vehicles, and each slave simply flies a reference primitive and returns a scalar cost. If all vehicles receive the same gains, their cost functions are averaged before the optimization step, which the experiments show reduces the scatter of the final gains under noisy altitude measurements; if each vehicle receives a different gain pair, one EQL iteration tests multiple points at once, which shortens the wall-clock tuning time. The method requires an even number of identical MAVs, is deterministic in duration because the number of iterations is fixed by the tolerance $\\epsilon$, and is demonstrated on the two altitude gains $k_P$ and $k_D$ of a geometric tracking controller, both in simulation and in outdoor flights with a pair of real hexarotors. The authors summarize the contribution as a reliable, scalable, time-deterministic, model-free procedure that improves tuning results while decreasing tuning time.","pith_inferences":["Beyond the paper: if the identical-platform assumption holds, the no-AVG speedup should scale roughly as $N$ for even $N$, since each EQL round tests $N$ gain pairs; the paper demonstrates only $N=2$, so the scaling law is my inference, not its claim.","Testable extension: the AVG noise-reduction claim implies that final-gain variance should fall as $1/N$ when the number of replicate vehicles grows, which could be checked directly by repeating the tuning campaign with $N=2,4,8$.","Testable extension: the same distribution logic should transfer to other zero-order line-search rules with the carry-over property, because the 40% saving comes from EQL's reuse of a tested point, not from the specific interval rule.","The flatness of the logged cost surface noted in the paper suggests that repeatability, not raw performance, is the main practical benefit of averaging; an operator might therefore prefer AVG for final refinement and no-AVG for initial exploration, a mixed protocol the paper does not itself propose."],"forward_implications":["With fixed tolerance $\\epsilon$ and known $K$, the total number of cost evaluations is known before takeoff, so a tuning flight can be planned inside battery endurance limits.","Averaging identical-gain cost readings across the swarm reduces the standard deviation of the final gains, as the simulation results show for both unloaded and payload-carrying platforms.","Testing different gain configurations on different vehicles reduces the tuning horizon, since one EQL round covers $N$ points of the search space instead of one.","For $N>2$, the two strategies can be combined: some vehicles replicate gains for averaging while others explore different gains, because the paper notes that results without AVG can still be averaged for $N>2$.","The staged, deterministic structure means the tuning procedure can be interrupted for battery changes or hardware fixes and resumed, leading toward a bank of controller gains for different payloads."],"supporting_citations":[{"why":"Supplies the original single-vehicle minimum-seeking tuning algorithm that this paper extends to a swarm.","marker":"[1]"},{"why":"Defines the AL-TUNE family and the EQL equal-division method; the fixed iteration count of EQL is what makes the parallel tuning time deterministic.","marker":"[2]"},{"why":"Demonstrates rapid single-MAV tuning that the parallel strategies build on and compare against.","marker":"[3]"},{"why":"Provides the rigid-body MAV model equations used by the controller and the simulation experiments.","marker":"[19]"},{"why":"Gives the SE(3) geometric tracking control law whose two altitude gains are the objects being tuned.","marker":"[20]"},{"why":"Supplies the MPC trajectory tracker that generates reference primitives and shapes the cost function surface.","marker":"[21]"},{"why":"Describes the UAV software stack used for state estimation, localization, and coordination in the real-world flights.","marker":"[22]"}],"fun_headline_variants":["Swarm tuning: average gains to cut noise, split gains to cut time","Parallel drone tuning trims noise and time by dividing or sharing gains","Rapid swarm tuning: distribute gain updates for less noise and faster results","Two gain-distribution tricks make drone swarm tuning rapid and robust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the vehicles in the swarm are effectively identical, so that for the same gains each vehicle's cost reading differs only by measurement noise and averaging cannot introduce bias; if one vehicle is heavier, less responsive, or sitting in a different wind field, the averaged or shared result can point the tuning toward the wrong gains.","fun_headline_variants_meta":{"raw":{"variants":["Swarm tuning: average gains to cut noise, split gains to cut time","Parallel drone tuning trims noise and time by dividing or sharing gains","Rapid swarm tuning: distribute gain updates for less noise and faster results","Two gain-distribution tricks make drone swarm tuning rapid and robust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2539,"prompt_tokens":887,"completion_tokens":1652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1575}},"tokens_in":503,"tokens_out":1652,"duration_ms":10739,"temperature":1.0,"reasoning_tokens":1575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:14:48.338018+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the AVG mode repeatedly with two vehicles and then with four vehicles under the same outdoor conditions; if the standard deviation of the final tuned gains does not decrease as the number of vehicles grows, the claimed noise-reduction benefit is not real. Alternatively, fly the no-AVG mode with two vehicles of measurably different mass and check whether the gain that is best for the vehicle that tested it is also best for the other vehicle; if not, parallel testing is trading correctness for speed.","supporting_citations":[{"cited_title":"Giernacki, D","cited_arxiv_id":null,"evidence_quote":"Supplies the original single-vehicle minimum-seeking tuning algorithm that this paper extends to a swarm."},{"cited_title":"Horla, W","cited_arxiv_id":null,"evidence_quote":"Defines the AL-TUNE family and the EQL equal-division method; the fixed iteration count of EQL is what makes the parallel tuning time deterministic."},{"cited_title":"Horla, M","cited_arxiv_id":null,"evidence_quote":"Demonstrates rapid single-MAV tuning that the parallel strategies build on and compare against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the rigid-body MAV model equations used by the controller and the simulation experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the SE(3) geometric tracking control law whose two altitude gains are the objects being tuned."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MPC trajectory tracker that generates reference primitives and shapes the cost function surface."}],"review_version":1}