{"id":"8b717ab6-6d4a-4efa-a0ef-e7d9e24fdb0a","arxiv_id":"2507.13903","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An aerial manipulator with trajectory optimization over a release interval, NMPC disturbance compensation, and online release timing reassessment reduces airdrop landing error to centimeter level in real flights.","lead":"An autonomous drone with a delta-arm manipulator plans and executes mid-air payload throws, using a release time window and predictive control to land objects within about 10 cm of a target, versus roughly 77 cm with fixed release timing in aggressive flights. The work shows that actively actuated throwing can absorb UAV tracking errors and release timing uncertainty better than passive dropping.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The release-timing reassessment that produces the headline improvement relies on the unquantified drag-free point-mass ballistic model; with a 2 cm cube released in the quadrotor wake, this model can bias the predicted landing-error sequence at the scale of the claimed precision.","rationale":"The paper presents a coherent integrated system with real flights and ablations: the continuous landing-point interval is a sensible mechanism for creating a release window, and the NMPC/NDOB/INDI ablation shows that the control stack improves landing-point concentration. The concern I find most load-bearing is the one the reader identified: the simplified projectile model of Sec. III-C is the basis for both the planner and the online release-timing reassessment, and its error is never quantified. This is not a generic 'model might be wrong' objection; it is specific to the decision variable that produces the Table II improvement. In the aggressive trajectory, the proposed reassessment reduces mean error from 76.7 cm to 10.1 cm, and the reassessment operates by comparing predicted landing points computed from the drag-free model. If aerodynamic drag, rotor downwash, spin, or residual electromagnet effects bias those predictions by even a few centimeters, the selected release time can be wrong in reality even though it appears optimal in the model. A free-flight calibration test with measured release states would directly measure this bias and settle whether the concern lands. Because the reader already made this the condition for acceptance, I do not change the verdict; the condition should be the delivery of that calibration or an equivalent quantitative error bound. Credit is due for the real-world comparative experiments and the ablation study, which are genuine evidence for the control-side contribution; the unquantified ballistic bias is the main gap separating the stated precision claim from the evidence.","tokens_in":11631,"tokens_out":11233,"duration_ms":142798,"concrete_test":"Fix a stationary rig carrying the same electromagnet and 2 cm iron cube, with optical ground-truth measurement of the release state and landing point, and release the payload from a set of measured initial states covering the experimental envelope (release speeds roughly 2.3–4.3 m/s and release heights/vertical velocities from the planned trajectories). For each shot, compare the measured landing position with the prediction from Eqs. (4)–(6) using the measured release state. Repeat with the payload released through the quadrotor wake (e.g., under a tethered quadrotor at nominal hover thrust) to isolate downwash effects. If the mean or 95th-percentile prediction error exceeds roughly 2 cm, the online reassessment's E(i) sequence in Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline precision is produced by the online release-timing reassessment (Alg. 1), which evaluates E(i) = ||Pl(i) − pW_t|| from predicted landing points computed via Eqs. (19a)–(19b) using the projectile model of Sec. III-C. That model assumes the payload is a point mass with no horizontal forces and vertical motion governed only by gravity after release. The released object is a 2 cm iron cube held by an electromagnet under a 1.59 kg quadrotor. At hover the rotors produce roughly 15.6 N of thrust; even a small fraction of the induced downwash crossing the cube produces aerodynamic forces on the order of 10^-2 N, and tumbling/spin plus residual magnetization can add further unmodeled effects. The paper explicitly acknowledges the simplification but provides no calibration measurement or error bound. Because the same biased ballistic predictor is used to choose the actual release time, a systematic model error does not average out: it shifts every predicted landing point in the horizon and can select a release state that is suboptimal in reality. The claimed 3–10 cm landing errors are the same order as plausible downwash-induced biases, so the central 'less sensitive to release timing' result is not yet separable from the unmodeled aerodynamics. The weakness is not that the model deviates from reality, but that the magnitude of the deviation is unquantified in exactly the quantity the system optimizes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an autonomous aerial throwing system based on a quadrotor-plus-delta-arm manipulator. The core ideas are: (i) a trajectory planner that imposes a smooth, time-windowed penalty on the predicted parabolic landing point, generating a feasible release interval rather than a single release instant; (ii) an NMPC controller augmented with NDOB and INDI hierarchical disturbance compensation; and (iii) an online release-timing reassessment algorithm that uses the NMPC prediction horizon to re-select the release time based on predicted landing errors. The paper reports simulation and real-world experiments, including an ablation of the control modules and comparative precision trials. The headline result is that, for an aggressive trajectory, the proposed reassessment reduces mean/max landing error from 76.7/83.8 cm to 10.1/12.3 cm.","tokens_in":11912,"tokens_out":4943,"duration_ms":64380,"significance":"If the central claims hold, the paper makes a useful contribution to autonomous airdrop: the interval-valued landing constraint is a sensible extension of point-wise landing constraints and seems to reduce sensitivity to release timing, while the NMPC-based reassessment is a practical way to exploit model-predictive information for triggering. The real-world validation on a physical aerial manipulator, including an ablation of NDOB and INDI, is valuable and goes beyond simulation-only studies. However, the strength of the experimental evidence is limited by missing statistical reporting, and the online reassessment depends on an unquantified ballistic-model assumption. The conceptual contribution is clear, but the load-bearing experimental and modeling evidence needs strengthening before the precision claims can be accepted at face value.","major_comments":[{"comment":"The central precision claims rest on MEAN/MAX (or RMSE/MAX) values with no trial counts, standard deviations, or confidence intervals. Figure 8 indicates ten throws for one trajectory, but Table II does not state n for each row, and Table I is similarly unspecified. With small sample sizes, the large improvement on the aggressive trajectory (76.7 to 10.1 cm) cannot be statistically distinguished from run-to-run variation. Please report per-trial landing errors, n, mean±SD, and confidence intervals or a suitable test for every condition.","section":"Section V-C, Table II and Section V-B, Table I"},{"comment":"The online release-timing reassessment selects the release instant using predicted landing points computed from a drag-free point-mass projectile model. The released object is a 2 cm iron cube, and the platform flies in its own rotor downwash; spin, residual magnetization, and aerodynamic forces can bias every predicted landing point. Because the same biased model selects the release time, the error does not average out across the prediction horizon. The claimed landing errors (3–10 cm) are the same order as plausible aerodynamic biases, yet the paper provides no calibration of predicted versus observed landing points and no sensitivity analysis. Please add a validation dataset of predicted vs. actual landing positions, or analyze how much the selected release time and final landing error change under bounded model perturbations.","section":"Section III-C and Section IV-C, Eqs. (19a)-(19b) and Algorithm 1"},{"comment":"The claim that the planned trajectories are 'less sensitive to release timing' is not quantitatively demonstrated. Figure 6 shows durations below unspecified error thresholds, but the text does not report the landing error as a function of release-time offset for different values of τ, or the relationship between τ, μ, and the achieved landing tolerance. To support the central contribution, please provide quantitative sensitivity curves, e.g., landing error versus release timing error for the optimized trajectories with and without the interval constraint.","section":"Section IV-A, Eqs. (8)-(10) and Section V-A"}],"minor_comments":[{"comment":"The symbol τ is used both for the release-window half-width in Eq. (9) and for the current time/loop variable in Algorithm 1, which is confusing; please use distinct symbols.","section":"Notation, Eq. (9) and Algorithm 1"},{"comment":"The projectile equations mix vector and scalar notation: g is defined as a scalar absolute gravitational acceleration but appears in vector expressions. Please define the gravitational acceleration vector and the landing-plane coordinate explicitly.","section":"Eqs. (4)-(6)"},{"comment":"There are several typographical issues, including 'UA V' with a stray space in the abstract and introduction; please proofread.","section":"Abstract and Introduction"},{"comment":"The caption refers to 'different threshold values' without stating the actual thresholds; please report these values and how they were chosen.","section":"Section V-A, Fig. 6"},{"comment":"The stopping condition 'if ∆t ≤ dt' appears to break whenever k* = 1; please clarify the indexing of k* and explain how actuator delay is explicitly incorporated into the decision rule.","section":"Section IV-C, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on two self-authored preprints ([12], [13]) for the delta-arm control and whole-body planning components; the incremental contribution over those works should be clarified for novelty assessment. The experimental evidence would also be considerably stronger if the authors made per-trial data and any launch scripts or model parameters available as supplementary material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a genuine systems contribution to autonomous aerial throwing. The new combination is the continuous release-window landing constraint (borrowed from tube acceleration work) applied to an aerial manipulator, plus an online release-timing reassessment using NMPC predictions. The real-world results are striking: in the aggressive trajectory, mean landing error drops from 76.7 cm to 10.1 cm. That's a real effect, not a simulation artifact.\n\nWhat's good: the planning formulation is clean, the relaxation function is well explained, and they show how the interval parameter changes the release action from throw to place. The hierarchical disturbance compensation (NDOB + INDI) is well motivated, and the ablation in Table I shows each piece earns its keep. The online reassessment is a clever reuse of the NMPC horizon.\n\nSoft spots, in order of seriousness. First, the projectile model. The planner and the reassessment both assume the 2 cm iron cube is a drag-free point mass after release. Under a hovering quadrotor, the downwash and tumbling can impose forces that plausibly bias predicted landing points by the same order as the claimed 10 cm precision. Because the same biased model selects the release time, a systematic error does not average out. The paper acknowledges the simplification but gives no calibration or error bound. This is the stress-test concern, and it holds up. It doesn't sink the main result—the sensitivity reduction from the continuous interval is geometric and model-independent—but it does mean the absolute precision numbers are only as good as the ballistic assumption.\n\nSecond, the statistics. Tables I and II report RMSE/MAX and MEAN/MAX, but no trial counts (except one figure caption mentioning ten throws), no standard deviations, no confidence intervals. For a claim this strong, that's a must.\n\nThird, no quantitative comparison against a point-wise constraint baseline. They cite prior work but don't show how much the interval helps relative to a single-instant constraint. Figure 6 is suggestive, not a head-to-head.\n\nFourth, no code or data. Not fatal, but this kind of integrated system is hard to reproduce without it.\n\nOverall, this is a decent paper that deserves a serious referee. The authors should fix the statistics and add a sensitivity analysis or calibration for the projectile model. If they do, it's a solid contribution. My recommendation: send it out, with those requests.\n\nRegards.","headline":"Solid aerial-throwing system paper with real experiments; the ballistic-model bias and missing error bars are fixable, but the core release-window idea is sound.","tokens_in":12481,"tokens_out":3986,"would_cite":true,"duration_ms":45987,"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":"A quadrotor carrying a delta arm can throw payloads to a target with roughly ten-centimeter accuracy even in aggressive flight, by planning a release-time window and re-timing the drop online from predictive states.","keywords":["aerial throwing","airdrop","aerial manipulator","trajectory optimization","nonlinear model predictive control","release timing reassessment","disturbance observer","payload delivery"],"falsifier":"Measure the actual trajectory of the 2 cm iron cube with a high-speed motion-capture system immediately after release and compare it with the ballistic prediction from Equations (4)-(6); if the observed flight path deviates systematically with release speed or payload orientation, or if wind-tunnel tests show landing bias that grows with crosswind speed, the point-mass assumption is falsified. A simpler check is to run the same planned trajectory with a payload of equal mass but larger cross-section and see whether the landing error grows, which it would if drag, rather than timing, dominates the residual error.","tokens_in":11398,"feed_emoji":"🎯","tokens_out":6865,"duration_ms":70566,"temperature":0.7,"pith_summary":"Autonomous airdrops from agile drones are hard because the exact instant of payload release strongly determines where the payload lands, and drones cannot track fast trajectories perfectly. This paper proposes a throwing system that makes the planned trajectory insensitive to release timing: instead of enforcing a landing constraint at a single release instant, the planner applies a smooth landing penalty over an entire time window around the release, so any release within that window still lands close to the target. The reference trajectory is tracked by a nonlinear model predictive controller with hierarchical disturbance compensation, and an online algorithm continuously re-evaluates the best release time from the controller's predicted states. In real flight tests the most aggressive trajectory landed within 10.1 cm mean error with the proposed method, versus 76.7 cm when releasing at the nominal time. If the approach scales beyond the tested platform, it would make fast, accurate aerial delivery practical without requiring the drone to hover or stop.","feed_headline":"Aerial robot arm throws payloads to within 10 cm","feed_subtitle":"Interval release windows plus predictive re-timing cut landing error from 77 cm to 10 cm.","key_machinery":"The central object is a smooth relaxation function $L_\\mu[x]$ that activates a landing-error penalty $G_l(t) = L_\\mu[E_t+\\tau]\\|P_l - p_t^W\\|^2$ over a time interval $\\tau$ around the nominal release time, rather than at a single instant. The interval is inserted into the MINCO-based trajectory optimization, so the output trajectory contains a continuous set of feasible release times $\\mathcal{T}_r = \\{t \\mid t \\in [t_r-\\tau, t_r+\\tau]\\}$. This is what converts a throw into a timing-tolerant task. The control stack then carries the argument: an NMPC over the quadrotor model tracks the reference, a nonlinear disturbance observer estimates the force change when the unknown-mass payload is dropped, an INDI inner loop rejects the configuration-dependent moment from the arm, and an online release-timing reassessment (Algorithm 1) queries the NMPC's predicted trajectory, evaluates the projectile landing point for every horizon state via Equations (19a)-(19b), and chooses the release instant with minimum predicted landing error.","core_discovery":"The core claim is that continuous, interval-based constraints on the parabolic landing point are the key to accurate aerial throwing. Point-wise landing constraints, as used in earlier work, create a trajectory that only works if release happens at one precise instant; any timing error translates directly into landing error. By smoothing the landing-penalty activation over a time window $\\tau$ around the optimal release time, the planner produces a feasible release interval $\\mathcal{T}_r$, and the optimizer exploits the spatial redundancy of the arm to slow or reposition the drone so release uncertainty has little effect. On the control side, a hierarchical disturbance compensation scheme, combining a nonlinear disturbance observer for linear forces with incremental nonlinear dynamic inversion for the payload-induced moment, is embedded in an NMPC framework. The NMPC's predictions are then used online to reassess the release timing: for every predicted state in the horizon the algorithm simulates the ballistic landing point and picks the instant that minimizes landing error. The paper reports that this combined system reduces landing error on the most aggressive tested trajectory from 76.7/83.8 cm (MEAN/MAX) with a nominal trigger to 10.1/12.3 cm with the proposed reassessment.","pith_inferences":["The interval-relaxation trick is generic: the same smoothed activation over a time window could make catching, perching, or ball-striking tasks robust to actuation delay, which is an extension the paper does not develop.","The centimeter-level numbers are tied to the point-mass ballistic model; for larger or non-spherical payloads, or in wind, aerodynamic drag would bias the predicted landing points, so the method would likely need a learned drag or wind model to keep its accuracy outdoors.","A direct test of the method's headroom is to replace the projectile prediction in Equations (19a)-(19b) with a learned residual map from release state to landing offset; if accuracy improves further, the residual error is model bias rather than timing, and if it does not, timing uncertainty is already the dominant term."],"forward_implications":["Drones can deliver payloads while flying fast, because the release window makes landing accuracy robust to timing jitter and actuator delay.","The same trajectory can be adjusted from a throw to a near place by increasing $\\tau$, giving mission planners a continuous trade-off between flight aggressiveness and precision.","The system tolerates unknown payload mass, since the disturbance observer and INDI loop absorb the model change at release without retuning.","Prediction-based release re-timing can correct for accumulated tracking errors in real time, which is the main reason aggressive trajectories improve from 76.7 cm to 10.1 cm error.","The approach narrows the gap between agile flight and delivery tasks, so airdrop no longer requires the drone to slow or hover over the target."],"supporting_citations":[{"why":"The point-wise parabolic constraint baseline that the proposed interval constraint is designed to improve.","marker":"[6]"},{"why":"Introduces the redundant throwing-configuration idea, adapted here to aerial throwing via a continuous release window.","marker":"[21]"},{"why":"Supplies the smoothed landing-penalty construction that the paper extends from a single instant to a time interval.","marker":"[26]"},{"why":"The MINCO trajectory optimization framework used to represent and minimize smooth polynomial spline trajectories.","marker":"[25]"},{"why":"Provides the unified corridor, kinematic, and dynamic constraint modeling for the whole-body planning problem.","marker":"[13]"},{"why":"Supplies the delta-arm control strategy and NDOB-based control used by the experimental aerial manipulator.","marker":"[12]"},{"why":"The INDI inner-loop law that compensates the payload-induced external moment during release.","marker":"[30]"},{"why":"The disturbance observer construction used to estimate the external force change when the payload is dropped.","marker":"[29]"}],"fun_headline_variants":["Interval release windows cut airdrop error to 10 cm","Predictive re-timing shrinks payload drop error to 10 cm","Aerial manipulator delivers 10-cm precise airdrops","Throwing drone hits 10-cm landing accuracy","Smooth release timing enables pinpoint airdrop accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The precision claim rests on the simplified projectile model: the payload is assumed to leave the end-effector with zero relative motion and then to fly under gravity alone, with no horizontal forces, so any drag, spin, or residual electromagnet effect would bias the predicted landing points and the release-time choice.","fun_headline_variants_meta":{"raw":{"variants":["Interval release windows cut airdrop error to 10 cm","Predictive re-timing shrinks payload drop error to 10 cm","Aerial manipulator delivers 10-cm precise airdrops","Throwing drone hits 10-cm landing accuracy","Smooth release timing enables pinpoint airdrop accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001204,"raw_usage":{"total_tokens":4966,"prompt_tokens":953,"completion_tokens":4013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":3928}},"tokens_in":569,"tokens_out":4013,"duration_ms":33331,"temperature":1.0,"reasoning_tokens":3928,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:13:37.247565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual trajectory of the 2 cm iron cube with a high-speed motion-capture system immediately after release and compare it with the ballistic prediction from Equations (4)-(6); if the observed flight path deviates systematically with release speed or payload orientation, or if wind-tunnel tests show landing bias that grows with crosswind speed, the point-mass assumption is falsified. A simpler check is to run the same planned trajectory with a payload of equal mass but larger cross-section and see whether the landing error grows, which it would if drag, rather than timing, dominates the residual error.","supporting_citations":[{"cited_title":"Fast trajectory optimization for agile quadrotor maneuvers with a cable- suspended payload,","cited_arxiv_id":null,"evidence_quote":"The point-wise parabolic constraint baseline that the proposed interval constraint is designed to improve."},{"cited_title":"Tube acceleration: robust dexterous throwing against release uncertainty,","cited_arxiv_id":null,"evidence_quote":"Introduces the redundant throwing-configuration idea, adapted here to aerial throwing via a continuous release window."},{"cited_title":"Real-time trajectory planning for aerial perching,","cited_arxiv_id":null,"evidence_quote":"Supplies the smoothed landing-penalty construction that the paper extends from a single instant to a time interval."},{"cited_title":"Geometrically constrained tra- jectory optimization for multicopters,","cited_arxiv_id":null,"evidence_quote":"The MINCO trajectory optimization framework used to represent and minimize smooth polynomial spline trajectories."},{"cited_title":"Whole-body integrated motion planning for aerial manipulators,","cited_arxiv_id":null,"evidence_quote":"Provides the unified corridor, kinematic, and dynamic constraint modeling for the whole-body planning problem."},{"cited_title":"Ndob-based control of a uav with delta-arm considering manipulator dynamics,","cited_arxiv_id":null,"evidence_quote":"Supplies the delta-arm control strategy and NDOB-based control used by the experimental aerial manipulator."},{"cited_title":"Accurate tracking of aggressive quadrotor tra- jectories using incremental nonlinear dynamic inversion and differential flatness,","cited_arxiv_id":null,"evidence_quote":"The INDI inner-loop law that compensates the payload-induced external moment during release."},{"cited_title":"Dob-based wind estimation of a uav using its onboard sensor,","cited_arxiv_id":null,"evidence_quote":"The disturbance observer construction used to estimate the external force change when the payload is dropped."}],"review_version":1}