{"id":"117f62d8-16ba-459d-b979-9b66423f6dc7","arxiv_id":"2505.03931","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"NMPC-Lander integrates NMPC with control barrier functions for autonomous drone landing on static and moving platforms, reporting 9 to 11 cm final point errors in real-world tests.","lead":"This paper combines nonlinear model predictive control with control barrier functions to land a drone on a static or moving platform while steering around static obstacles. It reports real-world landing errors of 9 to 11 centimeters and claims a nearly threefold accuracy improvement over a B-spline and A* baseline from earlier literature.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'nearly threefold' outperformance claim is unsupported: Table II compares a simulation-only B-spline+A* result from a different paper against this paper's real-world NMPC-Lander result under different conditions.","rationale":"The reader's conditional verdict already flags the inequitable comparison in its rationale, so my concern aligns with that part of the assessment. However, the reader's weakest_assumption focuses on the CBF safety guarantee, whereas I see the unmatched baseline comparison as the most load-bearing issue because it underpins the paper's headline claim of outperforming an existing planning method by nearly threefold. I also agree that the CBF constraint in Eq. 11 is not verified as a valid discrete-time CBF for this 12-state system; the relative-degree and feasibility conditions from Zeng et al. [25] are not checked, and h(x,y) appears to have relative degree two under the Euler-discretized dynamics. That is a serious secondary concern about the formal safety claim. I give credit for the real-world experiments, the 10-trial repetition, and the onboard computation, which support the internal precision numbers, but the lack of error bars and the internal average inconsistency (8.2 cm vs 8.7 cm for dynamic real-world) further weaken quantitative confidence. The verdict remains CONDITIONAL: the landing-precision claim can be accepted with caution, but the comparative and formal safety claims need additional evidence before full acceptance.","tokens_in":7779,"tokens_out":6592,"duration_ms":73418,"concrete_test":"Reproduce the comparison under identical conditions: implement or obtain the B-spline + A* planner from Guo et al. [18] and run it on the authors' real or simulated quadrotor with the same static obstacle, the same dynamic platform trajectory at the same speed, and the same VICON-based FPE metric, for at least 10 trials each. If the baseline mean FPE remains approximately three times larger than NMPC-Lander's 11 cm under matched conditions, the comparative claim holds; otherwise the 'nearly threefold' statement should be retracted or reworded to reflect a non-comparable literature benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section IV-C assert that NMPC-Lander outperforms the B-spline combined with A* planning method by nearly threefold in position tracking, citing 34.6 cm versus 11 cm (Table II). This comparison is not valid as stated. The 34.6 cm value is a simulation-only result taken from Guo et al. [18], with platform velocity 0.8 m/s and an unspecified obstacle and evaluation setup, while the 11 cm value is this paper's real-world dynamic-platform-with-obstacle result at 1.0 m/s. The two numbers come from different platforms, different perception and state-estimation pipelines, different obstacle layouts, and different environmental conditions (simulation versus real hardware with VICON, onboard Orange Pi 5B, and ground effects). No B-spline+A* baseline was run in the authors' own setup, and no matched comparison at the same platform velocity, with the same obstacle and the same FPE metric, is provided. Therefore the ratio 34.6/11 ≈ 3.1 does not establish a threefold improvement. The internal landing-precision claim (9.0 cm static, 11 cm dynamic in real-world obstacle scenarios) is plausible from the reported measurements, but the comparative claim is a headline contribution and it is insecure. This is the most load-bearing weakness because it directly affects the paper's central claim of superior robustness and practical effectiveness relative to an existing planning method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes NMPC-Lander, an onboard nonlinear model predictive control (NMPC) architecture combined with a control barrier function (CBF) constraint for autonomous quadrotor landing on static and moving platforms. The paper presents a 12-state quadrotor model with a ground-effect modification, an NMPC formulation with multiple shooting (Eqs. 4–7), an additional platform-tracking positional cost (Eq. 8), and a CBF constraint (Eq. 11) for static obstacle avoidance. Evaluation is carried out in Gazebo simulation and in real-world experiments with VICON motion capture and an Orange Pi 5B onboard computer. Reported average final point errors (FPE) from Table I are 2.1 cm (simulation, no obstacle) and 3.9 cm (real-world, no obstacle) for static landing, rising to 3.4 cm and 9.0 cm with an obstacle; for dynamic landing the corresponding values are 2.5 cm and 6.4 cm without obstacle, and 5.2 cm and 11.0 cm with obstacle. The paper claims that NMPC-Lander outperforms a B-spline combined with A* planning method by nearly threefold based on Table II, which compares an 11 cm real-world FPE with a 34.6 cm literature simulation value.","tokens_in":8101,"tokens_out":3157,"duration_ms":32009,"significance":"If the measured FPE values are reproducible, the paper provides a useful hardware demonstration of onboard NMPC-based landing with obstacle avoidance, and the simulated-versus-real consistency in Table I is a genuine strength. The contribution is a system integration result rather than a new theoretical one: the CBF guarantee is imported from [25] without new derivation, and the comparative claim is presented as a headline result. The measured landing-precision claims are self-contained and plausible, but the unsupported cross-paper comparison and the unverified safety preconditions are load-bearing issues that currently limit the paper's conclusions.","major_comments":[{"comment":"The claimed 'nearly threefold' improvement over the B-spline combined with A* method is not established by the data. The 34.6 cm value is a single literature simulation result from Guo et al. [18] with a platform velocity of 0.8 m/s, while the 11 cm value is the authors' real-world dynamic-platform result at 1.0 m/s, obtained with VICON state feedback, an Orange Pi 5B onboard pipeline, and a different obstacle layout and evaluation protocol. No B-spline + A* baseline was run in the authors' own setup, and the FPE metric is not matched across the two systems. Because this comparison appears in the abstract, the comparative-analysis section, and the conclusion, it is a load-bearing claim. It should be removed or replaced with a controlled baseline in the authors' own simulation and hardware settings; otherwise the statement of 'nearly threefold' outperformance is unsupported.","section":"Section IV-C, Table II"},{"comment":"The discrete-time CBF constraint is asserted to provide a formal safety guarantee by citing [25], but the preconditions of that theory are not verified. The barrier h(x,y) in Eq. (9) depends only on position, while the control inputs u enter through the acceleration channel of the 12-state dynamics, so the relative degree of h with respect to u is not established. The paper also does not define the discrete-time difference Δh in Eq. (11), does not justify the chosen γ=0.4, and does not show recursive feasibility of the CBF constraint over the prediction horizon. Consequently the statement that Eq. (11) 'ensures that the drone maintains a safe distance from the obstacle at each time step' exceeds what the manuscript demonstrates. The authors should either verify the discrete-time CBF conditions for the actual sampled dynamics or revise the claim to empirical obstacle avoidance based on the experiments.","section":"Section III-D, Eq. (11)"}],"minor_comments":[{"comment":"The text contains a typo: 'UA V' should be 'UAV'; the PDF also shows irregular spacing before punctuation in several places.","section":"Abstract"},{"comment":"The text reports an average FPE of 'approximately 6.5 cm' for real-world static landing and '8.2 cm' for real-world dynamic landing; these appear to be averages of the two real-world rows in Table I, but this should be stated explicitly to avoid apparent inconsistency with the individual FPE values of 9.0 cm and 11.0 cm.","section":"Section IV-B, Table I"},{"comment":"The ground-effect factor TIGE is written as a multiplicative reduction of the rotor thrust, whereas ground effect classically increases thrust near the ground; please clarify the sign convention and whether TIGE denotes an efficiency factor or the actual thrust, since Eq. (1) uses it as a thrust multiplier.","section":"Eq. (2)"},{"comment":"The CBF as defined in Eq. (9) handles a single point obstacle in the x-y plane. The paper does not discuss extension to multiple obstacles, obstacles with vertical extent, or obstacles whose position is time-varying; a sentence on these limitations would improve clarity.","section":"Section III-D"},{"comment":"The predicted-versus-observed plots appear to show a single representative run. Since Table I reports averages over 10 trials, reporting mean trajectories with standard deviations or an explicit error metric per trial would strengthen the experimental presentation.","section":"Figures 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"The core landing-precision measurements are plausible and the hardware demonstration is useful, but the comparative claim in Table II is a central advertised contribution and is currently unsupported. If the authors remove the cross-paper comparison or anchor it with a matched baseline, and if they temper the formal-safety claim to the level actually verified, the paper could be suitable for publication. The self-citations to the authors' own lab are frequent but appear topically relevant."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's own landing results are real and worth knowing. They report 9 cm static and 11 cm dynamic FPE in real-world obstacle landings, run fully onboard at 10 Hz with a 12-state NMPC plus discrete CBF. That is a practical, reproducible-looking engineering result. The Gazebo and real-world numbers are in a sensible range, and the ground-effect model is a nice touch.\n\nWhat is actually new is not the components—NMPC landing and CBF-MPC both exist—but the specific integration of a 12-state NMPC with a discrete-time CBF for moving-platform landing, demonstrated on hardware. I would not call it a new paradigm, but it is an honest engineering contribution.\n\nNow the soft spots, in order of severity. The abstract and Section IV-C claim the method outperforms B-spline+A* by 'nearly threefold.' That is not a controlled comparison. The 34.6 cm number is a single simulation result from Guo et al. with a slower platform (0.8 m/s), different perception, and no same-condition baseline. No B-spline+A* was run in this paper's setup. The ratio 34.6/11 does not establish the claim. This is the most load-bearing weakness because it is in the abstract and conclusion.\n\nSecond, the 'formal safety guarantee' of CBF is asserted from [25] without checking the discrete-time CBF conditions for this 12-state system. The barrier depends on position only, and the relative-degree/feasibility questions are not addressed. The experiments show obstacle avoidance in their scenarios, but the theory is not established. Soften the claim or verify it.\n\nThird, there are small internal inconsistencies: the dynamic real-world average is given as 8.2 cm but 6.4 and 11.0 average to 8.7. The static real-world 'average' 6.5 cm is presumably the mean of 3.9 and 9.0, but it is presented ambiguously after saying with-obstacle is 9.0. No error bars are reported for any of the 10-trial means.\n\nThe citation pattern looks fine; the self-citations are relevant. The math in the model is standard. The paper is not sloppy in its core derivation.\n\nWho is this for? Robotics practitioners working on autonomous landing and onboard MPC. They will get a useful data point and a sensible system design. Theory people will not. I would send it to a serious referee, but with a clear note that the comparative claim needs to be removed or re-run under matched conditions. As it stands, it is an honest engineering demo with one unsupported headline.","headline":"A useful onboard NMPC-CBF landing demo with solid FPE numbers, but the 'nearly threefold' improvement over B-spline+A* is not supported by the evidence.","tokens_in":8657,"tokens_out":3581,"would_cite":true,"duration_ms":30469,"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":"The paper claims that a single full-state nonlinear model predictive controller with a discrete-time control barrier function can land a quadrotor on static and moving platforms while avoiding obstacles, achieving average final errors of…","keywords":["Nonlinear Model Predictive Control","Control Barrier Function","Autonomous landing","Mobile landing platform","Quadrotor","Obstacle avoidance","Real-time onboard control"],"falsifier":"Repeat the obstacle-landing experiments with the obstacle placed closer than the 30 cm safety margin and record the minimum distance achieved across all trials; a single trial in which the drone enters the safety disk, or a single optimization failure during descent, would refute the claim that the barrier constraint enforces safety.","tokens_in":7552,"feed_emoji":"🚁","tokens_out":7269,"duration_ms":68089,"temperature":0.7,"pith_summary":"This paper tries to establish that one predictive controller, running entirely onboard, can both steer a quadrotor toward a landing platform and keep it away from obstacles, with no separate path planner. It embeds a positional safety constraint into a 12-state nonlinear model predictive controller and tests the result in simulation and on real hardware. The measured outcome is a final landing error of 9.0 cm on a static platform and 11 cm on a moving platform when an obstacle lies in the descent path, and the controller beats a B-spline-with-A* planner by roughly a factor of three in position tracking. If this holds, routine battery-swap and delivery landings could drop the need for external planning infrastructure.","feed_headline":"One controller lands drones on moving targets within 11 cm","feed_subtitle":"A single predictive controller with barrier safety constraints beats a planning-plus-tracking baseline by nearly threefold.","key_machinery":"The load-bearing mechanism is the control-barrier-function constraint appended to the NMPC optimization. The paper defines $h(x,y) = (x-x_{obs})^2 + (y-y_{obs})^2 - r_{safe}^2$, with $r_{safe} = r_{obs} + 30$ cm, and enforces $\\Delta h \\ge -\\gamma h$ at every prediction step, so predicted positions stay outside a safety disk around each obstacle. This constraint is carried by an NMPC formulated with multiple shooting over a horizon of ten steps at 0.1 s sampling, a full 12-state quadrotor model, a terminal cost on the final state, and a ground-effect term in vertical acceleration. The barrier function supplies the safety layer; the NMPC supplies tracking and landing.","core_discovery":"The central claim is that discrete-time Control Barrier Function constraints can be folded directly into a full-state NMPC formulation for landing, so a single optimization problem generates the descent trajectory, tracks the platform, and maintains a 30 cm safety margin around static obstacles. On real hardware with state feedback from a motion-capture system, the approach achieves average final point errors of 9.0 cm for static and 11 cm for dynamic platforms in obstacle scenarios; the dynamic-platform test used a target moving at 1 m/s. The same controller also outperforms the B-spline combined with A* method by nearly threefold in final position error, 11 cm versus 34.6 cm.","pith_inferences":["An extension the paper leaves implicit is to lift the barrier from horizontal position to the full 12-state vector, which would let the same controller avoid moving obstacles rather than only static disks.","Because the controller already tracks a moving platform, adding a prediction of platform motion would likely reduce the 11 cm dynamic error without changing the architecture.","The threefold comparison should be read carefully: the baseline ran in simulation at 0.8 m/s, while the proposed controller ran on hardware at 1.0 m/s; a matched hardware comparison would isolate the source of the gain."],"forward_implications":["A single onboard controller can replace the separate path planner and tracking controller in a landing stack.","Obstacle avoidance is enforced inside the optimization, so the descent trajectory does not need to be re-planned when a static obstacle appears.","Adding an obstacle increases real-world landing error by only a few centimeters: from 3.9 to 9.0 cm for static platforms and from 6.4 to 11 cm for dynamic platforms.","On a platform moving at 1 m/s, the controller keeps final landing error under 11 cm while maintaining a 30 cm safety margin around the obstacle."],"supporting_citations":[{"why":"Supplies the discrete-time control barrier function constraint used in Eq. (11), the formal safety layer of the architecture.","marker":"[25]"},{"why":"Provides the multiple-shooting NMPC formulation that the controller's optimization is built on.","marker":"[17]"},{"why":"The B-spline with A* method used as the experimental baseline for the comparison in Table II.","marker":"[18]"},{"why":"Ground-effect model incorporated into the vertical acceleration dynamics for landing.","marker":"[23]"},{"why":"The nonlinear optimization and symbolic-derivative framework used to solve the NMPC problem at each step.","marker":"[24]"}],"fun_headline_variants":["One predictive controller lands drone on moving target within 11 cm","NMPC with safety barrier lands drone on moving platform, error 11 cm","Controller beats A* by 3x: 11 cm error on moving landing, 9 cm static","Single NMPC lands drones on static and moving targets with 9-11 cm error","Drone landing: one controller, 30 cm safety margin, 11 cm moving accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's safety claim rests on the assumption that the discrete-time barrier constraint $\\Delta h \\ge -\\gamma h$ is satisfiable at every prediction step, but the barrier measures only horizontal position while the control inputs act through acceleration, and the relative-degree and feasibility conditions of barrier-function theory are never checked.","fun_headline_variants_meta":{"raw":{"variants":["One predictive controller lands drone on moving target within 11 cm","NMPC with safety barrier lands drone on moving platform, error 11 cm","Controller beats A* by 3x: 11 cm error on moving landing, 9 cm static","Single NMPC lands drones on static and moving targets with 9-11 cm error","Drone landing: one controller, 30 cm safety margin, 11 cm moving accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000689,"raw_usage":{"total_tokens":3078,"prompt_tokens":857,"completion_tokens":2221,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2110}},"tokens_in":473,"tokens_out":2221,"duration_ms":15215,"temperature":1.0,"reasoning_tokens":2110,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:41:45.853302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the obstacle-landing experiments with the obstacle placed closer than the 30 cm safety margin and record the minimum distance achieved across all trials; a single trial in which the drone enters the safety disk, or a single optimization failure during descent, would refute the claim that the barrier constraint enforces safety.","supporting_citations":[{"cited_title":"Safety-critical model predictive control with discrete-time control barrier function,","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete-time control barrier function constraint used in Eq. (11), the formal safety layer of the architecture."},{"cited_title":"Robust collision avoidance for multiple micro aerial vehicles using nonlinear model predictive control,","cited_arxiv_id":null,"evidence_quote":"Provides the multiple-shooting NMPC formulation that the controller's optimization is built on."},{"cited_title":"Simultaneous obstacles avoidance and robust autonomous landing of a uav on a moving vehicle,","cited_arxiv_id":null,"evidence_quote":"The B-spline with A* method used as the experimental baseline for the comparison in Table II."},{"cited_title":"Autonomous landing of quadrotor based on ground effect modelling,","cited_arxiv_id":null,"evidence_quote":"Ground-effect model incorporated into the vertical acceleration dynamics for landing."},{"cited_title":"CasADi: a software framework for nonlinear optimization and opti- mal control,","cited_arxiv_id":null,"evidence_quote":"The nonlinear optimization and symbolic-derivative framework used to solve the NMPC problem at each step."}],"review_version":1}