{"id":"3b389a6e-b26b-44fb-a745-98e4e1d18b0f","arxiv_id":"1908.03166","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding disturbance estimation to a soft-constrained MPC controller cuts peak position error during 10-12 m/s wind gusts by 71-78% versus a PID baseline, and an EKF matches a UKF at 2.5x lower compute.","lead":"This paper experimentally compares two quadrotor position controllers, MPC and PID, both augmented with Kalman-filter disturbance estimation to reject wind gusts and ground effect. It finds that a soft-constrained MPC with disturbance compensation reduces peak position error by 71-78% versus a PID baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I lacks repeated-run statistics for the headline 71-78% wind-gust improvement; without run counts and variance the central claim is not yet supported.","rationale":"The most load-bearing concern is empirical support for the quantitative headline claim. The paper is otherwise well constructed: the estimators are validated with step-response tests (Figs. 6-8), the computational cost measurements are concrete, and the EKF vs UKF comparison is practical and fair. However, the wind-gust results in Table I have no error bars. The authors even show in the ground-effect experiment that they are capable of reporting mean ± 2σ statistics (Fig. 11), so the omission in the wind-gust section is not a general lack of rigor. The turbulence of a fan-generated wind field makes the 'maximum deviation' a highly variable quantity. A single run serves as one sample; without knowing how many samples, the reported 71% and 78% improvements could be inflated by one good MPC run and one bad PID run. The review's CONDITIONAL verdict is therefore appropriate: the central claim is plausible but not yet adequately evidenced. Our attack is essentially identical to the reader's weakest assumption, so we agree. The concrete test—providing per-run data or re-running with at least 5 repetitions and computing confidence intervals—would settle the issue. If the CIs do not overlap, the claim stands; if they overlap, the claim should be weakened to a qualitative demonstration.","tokens_in":9749,"tokens_out":6831,"duration_ms":72793,"concrete_test":"Request the per-run maximum horizontal deviations for every entry of Table I, or re-run the 10 m/s and 12 m/s experiments at least five times per controller condition in randomized order, logging the anemometer time series to characterize the wind. Compute mean, standard deviation, and 95% confidence intervals for each controller; then test the differences (e.g., paired t-test or Wilcoxon signed-rank if the runs are paired by trial index). If the confidence interval for MPC+slack+comp overlaps PID+comp, or the sample size is below 5, the headline improvement over the PID baseline is not statistically established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that soft-constrained MPC with disturbance estimation cuts maximum horizontal deviation under 10-12 m/s wind gusts by 71-78% versus a PID baseline—rests on Table I, but that table reports only point maxima. Section V-F says 'repeated vertical landings' without specifying the number of runs, per-run values, or any variance or inferential statistic. The wind field is described as a 'high-powered fan' producing 'turbulent wind' (V-F-a), so the maximum deviation is a high-variance statistic: one favorable run for MPC+slack+comp and one unfavorable run for PID could produce a large ratio. The ground-effect experiments (V-E-b) report 'mean and 2σ bounds', demonstrating that the authors do compute statistics when it matters; their absence in the wind-gust experiments is conspicuous. Additionally, the tuning protocol ('overdamped position step response with minimum rise time', V-F-a) is too underspecified to verify that the PID and MPC are at comparable tuning levels. Without run counts and error bars, the 71-78% improvement is not statistically distinguishable from run-to-run luck.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an experimental comparison of PID and nonlinear MPC position controllers for a small quadrotor, both augmented with an Extended or Unscented Kalman Filter disturbance estimator. The authors model rotor thrust and drag maps with a voltage dependence, formulate an EKF and a UKF for external force and torque estimation, and integrate the resulting disturbance estimate as feedforward compensation in both PID and MPC architectures. They evaluate the system in three ways: ground-effect rejection during horizontal flights over a table edge, wind-gust rejection during vertical landings under 3, 10, and 12 m/s fan-generated wind, and computational cost on an Odroid XU4 companion computer. The main reported results are that the EKF achieves nearly the same estimation quality as the UKF at about 2.5x lower computational cost, and that a soft-constrained MPC with disturbance compensation reduces maximum horizontal deviation during strong wind gusts by 71-78% relative to an uncompensated PID baseline. The paper also includes an outdoor GPS-based validation of the EKF disturbance estimator and reports timing measurements for all controllers.","tokens_in":10002,"tokens_out":3724,"duration_ms":43789,"significance":"If the reported numbers are statistically sound, the paper provides useful experimental evidence on a question that is often addressed only in simulation: whether the added complexity of MPC and of UKF over EKF pays off on a computationally constrained multirotor. The strengths of the paper include the repeated ground-effect experiments with mean and 2-sigma bounds, the computational benchmarking on real hardware, the extension of disturbance estimation to GPS-grade state estimation outdoors, and the inclusion of a conventional PID baseline rather than only an MPC-versus-MPC comparison. The EKF/UKF comparison is empirical rather than derivation-based, and I see no circularity in the claims. The main weakness is that the central wind-gust result is reported as point maxima from an unspecified number of runs, with no variance or inferential statistics, which currently leaves the headline improvement unsupported.","major_comments":[{"comment":"The central quantitative claim, a 71-78% reduction in maximum horizontal deviation, is reported as a point maximum for each controller, with no run counts, per-run values, variance, or statistical test. Section V-F-a only says 'repeated vertical landings,' and the experiment is described as involving 'turbulent wind,' so the maximum deviation is a high-variance statistic. The ground-effect experiment in Section V-E-b reports mean and 2-sigma bounds, demonstrating that such statistics are computable on this platform; their absence in the wind-gust experiment is conspicuous. Please report the number of runs per condition, per-run deviations, mean plus/minus standard deviation (or median and range), and state whether the order of controllers was randomized or the wind conditions were otherwise matched across runs.","section":"Section V-F, Table I"},{"comment":"The claim that 'both the MPC and PID controllers were tuned to achieve an overdamped position step response with minimum rise time' is too underspecified to establish that the two controllers are at comparable tuning levels, which is important because the headline improvement is relative to a PID baseline. The tuning procedure should be described precisely: what step input was used, what rise time and overshoot were measured, how the gains were chosen, and whether the disturbance-estimator noise variance (identified as a tuning parameter in Section III-A-d) was set identically for all compensated controllers. Without this information, the comparison cannot be independently reproduced or assessed for fairness.","section":"Section V-F-a"},{"comment":"The metric underlying Table I is not unambiguously defined. The text says the controllers are compared by 'maximum deviations from the position reference,' Figures 12-14 show position error in the Z-X plane, and Table I reports 'Max. horizontal deviation.' Please define the metric precisely, including the coordinate frame, whether the maximum is over the two-dimensional horizontal error or the in-plane error shown in the figures, and whether it is taken over the full maneuver or only over the wind-affected segment. In addition, the wind speed is reported only as a single anemometer measurement in the center of the stream; because the deviation likely depends on the vehicle's distance from the fan and on turbulence, the paper should state the sensor placement relative to the flight path and, ideally, provide a wind profile or repeated wind-speed measurements.","section":"Section V-F-b and Table I"}],"minor_comments":[{"comment":"There is a typo in 'whethe the choice of MPC' that should be corrected.","section":"Section I-B"},{"comment":"The word 'preferrable' should be 'preferable.'","section":"Section III"},{"comment":"The word 'repeatibility' should be 'repeatability.'","section":"Section V-B"},{"comment":"The statement that without disturbance compensation 'both the PID and the MPC controllers experience an identical deviation' is made without supporting repeated-run data; adding the per-controller traces or statistics would make this claim verifiable.","section":"Section V-F-b"},{"comment":"The figures would be clearer if the wind-affected region and the maximum deviation used in Table I were marked explicitly on the position-error plots.","section":"Figures 12-14"}],"recommendation":"major_revision","confidential_remarks":"The experimental design is thoughtful, but the central comparison needs repeated-run statistics and a fully specified tuning protocol before the paper can be accepted. I would also encourage the authors to make the dataset or at least per-run traces available, since the 'extensive' claim is currently difficult to verify from the text alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Roughly: this is a genuinely useful experimental comparison, not a breakthrough. It does something prior work hadn't: put MPC and PID head-to-head on the same airframe with active disturbance estimation, and it also gives a clean EKF-vs-UKF comparison indoors and outdoors. The ground-effect results (mean and 2σ bounds) and the computational-cost measurements are the most solid parts.\n\nThe soft spot is exactly where the stress-test points. Table I reports single maximum horizontal deviations for each controller and wind speed, with no run counts, no variance, no statistical test. The text says 'repeated vertical landings' but never says how many or shows per-run values. Given that the wind field from a fan is turbulent, the max deviation is a high-variance statistic, so the headline 71–78% improvement over PID is not yet supported. It could be real—the pattern across 10 and 12 m/s and across the controller variants is consistent—but the numbers should not be quoted without error bars.\n\nThe tuning protocol is also underspecified. 'Overdamped position step response with minimum rise time' is not enough to verify that PID and MPC were at comparable tuning levels. The authors should specify the tuning procedure or provide the resulting gains and weights.\n\nMinor: the EKF-vs-UKF conclusion rests on a visual comparison of one step-response plot (Figure 7) plus computational cost. That's probably fine for an engineering paper, but a more quantitative noise/error analysis would strengthen it.\n\nNo code or data is released, which is unfortunate for reproducibility, but the paper gives enough architectural detail that others could reimplement.\n\nOverall: this is a solid, useful engineering paper. The qualitative conclusion—disturbance compensation helps a lot, and soft-constrained MPC gives the best rejection—is convincing. The quantitative claims need support. A serious referee should ask for run counts, error bars, and tuning details before publication. I'd send it to review, not desk-reject.","headline":"A useful experimental comparison whose headline wind-gust numbers need run-to-run statistics before you trust them.","tokens_in":10529,"tokens_out":2590,"would_cite":true,"duration_ms":26611,"reading_group":"yes","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 an EKF disturbance estimate plus a soft-constrained nonlinear MPC cuts maximum horizontal position drift in 10-12 m/s wind gusts by 71-78% versus a PID baseline, with the EKF matching a UKF at 2.5x lower cost.","keywords":["multirotor position control","disturbance estimation","extended Kalman filter","unscented Kalman filter","model predictive control","wind gust rejection","ground effect"],"falsifier":"Repeat the 10 and 12 m/s gust landing experiments many times with randomized entry states and a measured wind field, then compare the distributions of maximum horizontal deviation; if the confidence intervals for the soft-constrained MPC with compensation and the PID with compensation overlap, the claimed 71-78% improvement over the PID baseline does not hold.","tokens_in":9563,"feed_emoji":"🌀","tokens_out":10245,"duration_ms":98021,"temperature":0.7,"pith_summary":"The paper asks whether the extra computational complexity of model predictive control (MPC) over a classical PID controller is worth it for high-precision multirotor position control in strong disturbances. It claims that adding an online disturbance estimate to either controller dramatically improves wind-gust rejection, and that the best combination—a soft-constrained MPC with an Extended Kalman Filter (EKF) disturbance estimate—cuts maximum horizontal position deviation by 71-78% in 10-12 m/s gusts relative to the PID baseline. It further claims that the simpler EKF matches an Unscented Kalman Filter (UKF) in estimate quality while running 2.5 times faster, so the EKF is the practical choice on a computationally constrained small drone. If these results hold, precision tasks such as docking, package delivery, and landing can stay accurate in heavy wind without requiring a large onboard computer.","feed_headline":"Wind-aware drone control cuts drift by up to 78%","feed_subtitle":"An onboard filter and soft-constrained MPC hold a small quadrotor to 5-10 cm in 10-12 m/s gusts.","key_machinery":"The central object is the disturbance-augmented estimation loop. An EKF (and, for comparison, a UKF) appends the external force vector $f_{\\mathrm{ext}}$ and external torque vector $\\eta_{\\mathrm{ext}}$ to the quadrotor state, models their time evolution as a Gaussian random walk, and corrects them with the onboard state estimator's position, velocity, and attitude measurements. The resulting force estimate enters the position controller as a feedforward acceleration term, while the MPC also uses the estimated disturbance inside its prediction model. The control law that carries the best results in the paper is the soft-constrained nonlinear MPC: a receding-horizon optimal control problem with a 2 s horizon, thrust and roll/pitch input constraints, a slack variable $\\epsilon$ that relaxes the cross-track error bound, and path-following guidance that steers the vehicle to the closest point on the reference path.","core_discovery":"On the paper's own terms, the central discovery is that disturbance rejection, not the controller class, dominates performance once wind gusts become strong. At 3 m/s, PID with disturbance compensation and MPC with disturbance compensation both nearly eliminate the position error and no clear difference appears between them. At 10 and 12 m/s, uncompensated MPC performs about the same as uncompensated PID (19 vs 17 cm and 39 vs 45 cm maximum horizontal deviation), while the EKF feedforward improves both; the soft-constrained MPC with the slack cross-track constraint and disturbance compensation reaches 5 cm and 10 cm, respectively, which is a 71% and 78% reduction from the PID baseline's 17 cm and 45 cm. The paper also reports that the EKF and UKF produce nearly identical disturbance estimates, with the EKF averaging 8 ms per iteration versus 20 ms for the UKF, and that the EKF remains usable outdoors with degraded GPS-based state estimation, though the torque estimate becomes noisier. In the ground-effect experiment, the PID with disturbance compensation stayed within 2.5 cm of the reference, whereas without compensation the vehicle overshot and undershot by 5-10 cm and showed integrator wind-up.","pith_inferences":["A testable extension the authors do not run is to replace the single-gust protocol with many randomized passes through a measured wind field and report per-run maxima; this would show whether the 71-78% margin is stable under run-to-run variability.","Because the torque estimate is not used in control, the paper leaves implicit that feeding estimated disturbance torque into the attitude loop might further reduce the residual position error during gust onset; this is an empirical question the experiments do not answer.","The paper's explanation that estimator convergence time limits rejection suggests an adaptive process-noise schedule—raising the random-walk variance right after a detected gust—could shorten the 1.1 s rise time without inducing roll/pitch oscillations.","The MPC treats the disturbance as constant over its 2 s horizon; extending the estimator to a time-varying or learned disturbance model inside the horizon could shrink the remaining 5-10 cm residual, though it would add computation the paper is explicitly trying to avoid."],"forward_implications":["Precision maneuvers such as docking, perching, and precision landing can tolerate 10-12 m/s gusts while holding horizontal deviation around 5-10 cm, instead of the 17-45 cm of an uncompensated PID controller.","On small, payload-limited drones, the EKF is a better choice than the UKF: it runs at 100 Hz versus 40 Hz and gives essentially the same force and torque estimates.","Even without MPC, a PID controller with a disturbance feedforward term gets most of the benefit in moderate gusts, so the added complexity of MPC pays off mainly in extreme conditions or when a hazard corridor must be respected.","The soft-constrained MPC's slack variable gives a practical way to enforce an accuracy bound on cross-track error without risking infeasible optimization problems."],"supporting_citations":[{"why":"Supplies the nonlinear MPC formulation and the EKF disturbance estimate this work adapts to path following and soft constraints.","marker":"[4]"},{"why":"Supplies the quaternion/MRP unscented filtering machinery used in the UKF comparator.","marker":"[14]"},{"why":"Frames the stochastic external force and torque estimation problem the filters are designed for.","marker":"[13]"},{"why":"Supplies the calibrated rotor thrust and torque map used as the filter's input model.","marker":"[15]"},{"why":"Supplies the code-generation toolkit that turns the optimal control problem into fast C++ code.","marker":"[21]"},{"why":"Supplies the quadratic programming solver used inside the MPC iteration.","marker":"[22]"},{"why":"Supplies the standard EKF and UKF prediction and update equations both filters use.","marker":"[17]"}],"fun_headline_variants":["Disturbance rejection, not controller choice, rules wind-affected drone flight","Onboard filter, not controller type, determines gusty drone performance","MPC with EKF holds drone to 10 cm under 12 m/s wind gusts","PID plus disturbance estimator reduces ground-effect drift to 2.5 cm","EKF matches UKF disturbance estimates at half the compute load"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 71-78% improvement is computed from maximum deviations over repeated landings, but the paper reports no run count, variance, or statistical test and only a qualitative manual tuning rule for both controllers, so if wind or entry conditions differed between runs the ranking could change.","fun_headline_variants_meta":{"raw":{"variants":["Disturbance rejection, not controller choice, rules wind-affected drone flight","Onboard filter, not controller type, determines gusty drone performance","MPC with EKF holds drone to 10 cm under 12 m/s wind gusts","PID plus disturbance estimator reduces ground-effect drift to 2.5 cm","EKF matches UKF disturbance estimates at half the compute load"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001541,"raw_usage":{"total_tokens":6168,"prompt_tokens":955,"completion_tokens":5213,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":5115}},"tokens_in":571,"tokens_out":5213,"duration_ms":40175,"temperature":1.0,"reasoning_tokens":5115,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:21:16.839242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the 10 and 12 m/s gust landing experiments many times with randomized entry states and a measured wind field, then compare the distributions of maximum horizontal deviation; if the confidence intervals for the soft-constrained MPC with compensation and the PID with compensation overlap, the claimed 71-78% improvement over the PID baseline does not hold.","supporting_citations":[{"cited_title":"Linear vs Nonlinear MPC for Trajectory Tracking Applied to Rotary Wing Micro Aerial Vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear MPC formulation and the EKF disturbance estimate this work adapts to path following and soft constraints."},{"cited_title":"Collaborative transportation using MA Vs via passive force control,","cited_arxiv_id":null,"evidence_quote":"Supplies the quaternion/MRP unscented filtering machinery used in the UKF comparator."},{"cited_title":"Unscented external force and torque estimation for quadrotors,","cited_arxiv_id":null,"evidence_quote":"Frames the stochastic external force and torque estimation problem the filters are designed for."},{"cited_title":"Thrust Mixing , Saturation , and Body-Rate Control for Accurate Aggressive Quadrotor Flight,","cited_arxiv_id":null,"evidence_quote":"Supplies the calibrated rotor thrust and torque map used as the filter's input model."},{"cited_title":"{ACADO} { T}oolkit – {A}n {O}pen {S}ource {F}ramework for {A}utomatic {C}ontrol and {D}ynamic {O}ptimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the code-generation toolkit that turns the optimal control problem into fast C++ code."},{"cited_title":"qpOASES: a parametric active-set algorithm for quadratic program- ming,","cited_arxiv_id":null,"evidence_quote":"Supplies the quadratic programming solver used inside the MPC iteration."},{"cited_title":"Simon, Optimal State Estimation: Kalman, H Infinity, and Nonlin- ear Approaches","cited_arxiv_id":null,"evidence_quote":"Supplies the standard EKF and UKF prediction and update equations both filters use."}],"review_version":1}