{"id":"3638b25c-282b-4450-9488-ec1f9665a317","arxiv_id":"2607.23867","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Battery- and propulsion-aware NMPC with online thrust-limit replanning cuts tracking RMSE sixfold and roughly doubles safe flight time versus uncompensated control in simulated obstacle courses.","lead":"A quadrotor controller predicts how battery drain shrinks available thrust in real time and replans the path so the vehicle stays flyable longer. That matters for drone racing and other high-speed flights where fixed thrust limits cause crashes as the pack empties.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline 100%/46%/6x gains are earned mostly in a regime where the battery model is pure cubic extrapolation: the sim runs SOC negative (§VII), outside the 300–3300 mAh range where eqs. 5–7 were fitted and validated (§VI), and T_max — hence flight-termination time — is directly set by those same","rationale":"The reader's weakest_assumption already bundled the sim-only headline gains, the raised 1.54 kg mass, and the negative-SOC discharge; I agree with that identification and am sharpening it to the specific causal channel: eqs. 6–7 cubic extrapolation feeds eqs. 15–16, which set T_max, which sets flight-termination/collision times — the exact quantities in the headline percentages. The reader's CONDITIONAL verdict already prices this in (\"accept-shaped if claims are scoped to 'model validated outdoors; replanning gains in stressed sim'\"), so my read does not move the verdict; it strengthens the rationale for requiring the rescoping and suggests a cheap, decisive check. I do not push to REJECT because the real-world contribution — real-time electrical prediction validated in agile flight at 100 Hz with 5 ms mean solve time, plus a working bench validation of T_max at full throttle (§V, Fig. 8) — stands independently of the sim packaging, and the sim itself is a legitimate demonstration that thrust-aware replanning dominates non-replanning qualitatively (Fig. 17a torque headroom is physically sensible). The secondary oddity that the thrust-aware-no-replan case collides earlier than the unaware case (Table III: 113.4 s vs 117.2 s) is unexplained and worth an erratum, but it is not load-bearing for the central claim. No code/data release limits independent verification, which the reader already scored.","tokens_in":20994,"tokens_out":3134,"duration_ms":79732,"concrete_test":"Two-step check. (a) Evaluate v_oc(soc) and r_b(soc) from the Table II coefficients over soc in [−1500, 300] mAh and compare against the manufacturer's 4S 3300 mAh discharge curves (or a bench discharge to cutoff): if extrapolated v_oc sits >0.5 V above the real knee or r_b fails to rise steeply near empty, T_max in late flight is inflated. (b) Re-run the three §VII experiments at the true 1.2 kg mass with flight terminated at soc = 0 (no negative extrapolation). If the replanning advantage drops well below the reported +100%/+46% (e.g., to <+30% time), the headline numbers are extrapolation artifacts and the abstract must be rescoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (collision-free flight, 6x RMSE reduction, +46% distance, +100% time) rests entirely on the §VII simulation, and within that simulation the differentiating behavior occurs late in the flight, after SOC falls below the validated floor (~300 mAh in §VI) and goes negative by design. This is load-bearing through a specific mechanism: v_oc(soc) and r_b(soc) are cubic polynomials (eqs. 6–7, Table II) fitted on discharge data over the physical capacity range. Cubics extrapolated past the fit domain are unconstrained; plugging the Table II coefficients at soc = −1000 mAh gives v_oc ≈ 12.9 V and r_b ≈ 0.013 Ohm — a smooth, gentle decay. A real 4S LiPo near and below empty shows a sharp voltage knee and a steep internal-resistance rise. Because T_max is computed from v_oc and r_b via eqs. 15–16, the extrapolated gentle decay keeps predicted T_max higher for longer than physics would, which directly extends the replanning case's flight time (234 s) and distance (2.96 km) — the numerators of the +100%/+46% claims — and shifts the baseline collision times (117 s/131 s) that form the denominators. Compounding this, the mass was raised to 1.54 kg so hover thrust (~15 N) and the 52.8 N required-thrust threshold sit close to the ~65 N ceiling, making collision/termination times highly sensitive to small T_max errors; a ±1–2 N error in extrapolated T_max can move flight-end times by tens of seconds. The 17% current error acknowledged in §VI-C was measured in the validated regime; nothing bounds the error where the headline numbers are generated. So the abstract's percentages may substantially be artifacts of polynomial extrapolation plus a stress-tuned mass, not properties of the method. The real-flight validation (v_oc convergence, r_b with Kt, 5 ms mean solve) does support the narrower claim that the model predicts electrical quantities adequately in-flight — that part is solid.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper presents BC-NMPC, a quadrotor NMPC formulation that integrates multivariate polynomial models of the motor-propeller unit (thrust/power/resistance as functions of throttle, voltage, RPM; eqs. 1-4) and of a 4S LiPo battery (IR model with v_oc(soc), r_b(soc) cubics; eqs. 5-7) to predict, in real time, terminal voltage, current, power, and maximum available collective thrust T_max (eqs. 15-16). T_max enters the OCP as a nonlinear constraint (eq. 17), and a PMM-based planner replans the trajectory in flight as the thrust limit shrinks. The component models are validated on a static full-airframe bench test (Fig. 8) and in outdoor lemniscate flights at up to 23 m/s and 44 m/s^2 down to ~300 mAh SOC (§VI), with post-landing v_oc consistency, a temperature-compensated r_b, terminal-voltage MAE of 0.20 V, ~17% current error, and mean NMPC solve time of 5 ms. The headline performance claims — collision-free flight, 6-fold RMSE reduction, +46% distance, +100% flight time versus uncompensated flight — come exclusively from a simulation study (§VII, Table III) in which the aircraft mass is increased to 1.54 kg and the battery is allowed to discharge into negative SOC.","tokens_in":21457,"tokens_out":3630,"duration_ms":83988,"significance":"If the claims hold, the contribution is meaningful for agile UAV flight: a closed-form propulsion/battery model fast enough to run inside NMPC at 100 Hz (mean solve time 5 ms, Fig. 14), validated in genuinely aggressive real flight (23 m/s, 44 m/s^2), with in-flight replanning against evolving thrust limits. The identification-then-validation methodology (bench identification, separate flight telemetry validation) is sound, and the paper is refreshingly explicit about its error budget. However, the headline quantitative gains rest entirely on a simulation whose decisive regime (negative SOC, increased mass) lies outside the validated model domain, so the significance of the numerical claims is currently overstated relative to what is demonstrated.","major_comments":[{"comment":"The battery polynomials v_oc(soc) (eq. 7) and r_b(soc) (eq. 6) are cubics fitted on discharge data over roughly 300-3300 mAh (Fig. 7, §VI), yet §VII explicitly 'allowed [the battery] to discharge into negative SOC.' Cubics extrapolated past the fit domain are unconstrained; evaluating the Table II coefficients at soc = -1000 mAh gives v_oc ≈ 12.9 V and r_b ≈ 0.013 Ohm — a gentle, smooth decay. A real 4S LiPo near and below empty exhibits a sharp voltage knee and a steep internal-resistance rise, neither of which a cubic can represent. Since T_max is computed from v_oc and r_b via eqs. (15)-(16), the extrapolation keeps predicted T_max artificially high late in flight. This directly sets the replanning case's flight time (234.4 s) and distance (2.96 km) — the numerators of the headline '+100% flight time / +46% distance' claims — and also shifts the baseline collision times (117.2 s / 131","section":"§VII, eqs. (6)-(7), (15)-(16), Table II"},{"comment":"The digital twin's mass was raised from the real 1.2 kg to 1.54 kg 'to highlight the performance degradation.' This places hover thrust (~15 N) and the 3.5 g Required Thrust Threshold (52.8 N) close to the ~65 N ceiling, so collision/termination times become highly sensitive to small T_max errors. Given the acknowledged 15-17% current-estimation error (§VI-C) and the extrapolated T_max of Major Comment 1, an error of 1-2 N in predicted T_max can plausibly move the flight-end and collision times in Table III by tens of seconds. The headline gains are therefore specific to a configuration that is both non-physical (heavier than the real aircraft) and operating in an unvalidated battery regime. At minimum, a sensitivity analysis over mass/RTT margin, or a rerun at the real mass, is needed to show the qualitative conclusion (replanning prevents collision) is robust rather than an artifact of","section":"§VII, Table III"},{"comment":"The abstract and conclusion state 'a 6-fold decrease in tracking RMSE, a 46% increase in flight distance, and a 100% increase in flight time in an obstacle-ridden environment' without qualification. Per the manuscript itself, these numbers come exclusively from the §VII simulation with increased mass and negative-SOC extrapolation; the real-world experiments (§VI) validate only the component predictions (v_oc, r_b, i, v_cc, solve time), and no real flight demonstrates the replanning benefit or a collision-avoidance outcome. The claims should be explicitly scoped as simulation results obtained under a modified configuration, and the abstract should not present them as demonstrated flight performance.","section":"Abstract; §VIII"},{"comment":"The nonlinear constraint (17) enforces sum_i f_i <= T_max,k using the raw predicted T_max with no safety margin, while §VI-C acknowledges ~17% current error (8.72 A MAE on a 50.71 A mean) attributed to unmodeled advance-ratio torque, sensor scaling, and battery recovery. Since T_max derives from the same v_oc/r_b/p_in chain, a comparable relative error on T_max is plausible; the manuscript never quantifies the T_max prediction error in flight (Fig. 8 validates T_max only at 100% throttle on the bench, i.e., static conditions). For a controller whose stated purpose is to fly 'at the constantly-changing platform limits,' the error budget of the limit itself should be characterized in dynamic flight, or a margin/robustification of (17) discussed.","section":"§IV-B, eq. (17); §VI-C"}],"minor_comments":[{"comment":"Numerous typos: 'demonstrates achieves' (abstract), 'dicuss' (§III-A), 'Therfore' (§IV-A), 'therfore' (§IV-B), 'mistmatch' (§VI-B), 'betwen' (Fig. 3 caption), 'examplesof' (§VI-E), 'upto' (multiple).","section":"Throughout"},{"comment":"Eq. (8): soc_dot = -i mixes units — soc is in mAh (Figs. 7, 10) while i is in A; the required 1000/3600 conversion factor should be stated for reproducibility of the RK4 integration.","section":"§IV-A, eq. (8)"},{"comment":"Eq. (15) uses r_m for the parallel combination of the four motor resistances, while eq. (3) and Fig. 2 define r_m per motor; the notation should distinguish r_m,parallel from r_m,i.","section":"§IV-B, eq. (15)"},{"comment":"§VI-B: the temperature-compensation coefficient K_t (eq. 19) is obtained from 'a short calibration flight,' but it is not stated whether battery temperature is measured in flight, how K_t generalizes across ambient temperatures, or whether the validation flights of Fig. 11b are independent of the calibration flight. Please clarify.","section":"§VI-B, eq. (19)"},{"comment":"Fig. 11 legends appear garbled ('r r̂_b'); the estimated vs. measured curves are hard to distinguish. Fig. 10: the legend for the two traces is missing/unclear.","section":"Figs. 10-11"},{"comment":"Eq. (16): writing T_max = 4(a_t v_cc^2 + b_t v_cc + d_t v_cc + c_t + e_t) would be clearer as the direct th=1 substitution of eq. (4), i.e., 4(a_t v_cc^2 + (b_t + d_t) v_cc + (c_t + e_t)); please check the printed grouping.","section":"§IV-B, eq. (16)"},{"comment":"Reference [20] is listed as 2024 but the arXiv identifier is from 2021; also verify the claim in §II-C ('to the best of our knowledge, the first approach...') against recent literature, since it is a strong novelty statement.","section":"References; §II-C"}],"recommendation":"major_revision","confidential_remarks":"The engineering contribution is real and the experimental work is solid, but the abstract's quantitative claims are drawn entirely from a simulation whose decisive regime (negative SOC, increased mass) is outside the validated model domain. This is fixable within the manuscript's scope — a rerun within the validated SOC range plus a sensitivity analysis would resolve it — but the authors should be pressed to either redo the simulation or clearly re-scope the headline numbers. The work appears to be a competent extension of the group's prior NMPC line ([33]) and fits the journal's scope well."},"author_rebuttal":{"model":"moonshotai/kimi-k3","summary":"We thank the referee for a careful and technically precise report. We agree with the central observation: the headline quantitative gains (6-fold RMSE, +46% distance, +100% flight time) are produced in a simulation that deliberately exaggerates the degradation regime by increasing mass to 1.54 kg and by allowing SOC to go negative, i.e., outside the domain in which the battery polynomials were identified (roughly 300-3300 mAh). We will revise to (i) scope all headline numbers explicitly as simulation results under a modified configuration, (ii) add a sensitivity study over mass/RTT margin and a rerun at the real 1.2 kg mass with SOC clamped at the validated lower bound, and (iii) characterize the T_max prediction error and discuss a safety margin on constraint (17). We maintain that the component-model validation (23 m/s, 44 m/s^2 real flight, 0.20 V terminal-voltage MAE, 5 ms solve time) stands on its own within the validated domain; what requires correction is the presentation and robustness evidence for the system-level claims, not the modeling methodology itself.","responses":[{"response":"We agree this is the decisive weakness of the simulation study, and we thank the referee for the concrete coefficient evaluation at soc = -1000 mAh, which makes the issue unambiguous. The negative-SOC discharge was intended as a qualitative stress test ('exaggerate the effects... for heavy aircraft and low-discharge chemistries', §VII), but the manuscript does not make clear that the resulting numbers inherit an unphysical smooth decay. In the revision we will: (1) re-run the §VII experiments with the SOC evolution clamped at the validated lower bound (~300 mAh) and a hard termination when v_oc(soc) reaches the fit-domain minimum, reporting flight time/distance/RMSE for all three cases under the clamped model; (2) add a supplementary run with a phenomenological knee model (e.g., sharp v_oc drop and r_b rise below the fit domain) to show how the replanning case behaves when the extrapolation is pessimistic rather than generous. We expect the qualitative ordering (replanning collision-free; baselines collide) to survive because the replanner's benefit accrues while SOC is still well inside the validated range, but the exact '+100% flight time' figure will change and will be reported as a clamped-model number with the knee-model result as a robustness bound.","revision_made":"yes","referee_comment":"Battery cubics v_oc(soc), r_b(soc) fitted over 300-3300 mAh are extrapolated into negative SOC in §VII, where a real 4S LiPo has a sharp voltage knee and steep r_b rise. This keeps predicted T_max artificially high and directly sets the 234.4 s / 2.96 km headline numbers."},{"response":"We agree a sensitivity analysis is necessary and will add it. The mass increase was chosen to make the degradation observable within a short simulated flight, but the manuscript presents the resulting numbers without acknowledging that they sit in a knife-edge regime. In the revision we will: (1) rerun all three cases at the real 1.2 kg mass with the clamped battery model; (2) sweep mass in, e.g., 0.1 kg steps between 1.2 and 1.54 kg and report collision time, flight time, distance, and mean RMSE for each of the three configurations; (3) report the margin between RTT and T_max over time for each run, so the reader can see where each case operates relative to the ceiling. This will demonstrate whether the qualitative conclusion (replanning prevents collision and extends safe flight) holds across the margin range or only near the 1.54 kg configuration, and we will state the outcome honestly either way.","revision_made":"yes","referee_comment":"The digital twin mass was raised to 1.54 kg, making hover thrust and the 52.8 N RTT sit close to the ~65 N ceiling, so Table III outcomes are highly sensitive to small T_max errors given the acknowledged 15-17% current error. A sensitivity analysis over mass/RTT margin or a rerun at real mass is needed."},{"response":"This is correct and we will fix it. The abstract will be rewritten to state explicitly that model accuracy was validated in real-world flight experiments while the performance gains were demonstrated in simulation under a modified (heavier) configuration, e.g.: 'The accuracy of the proposed model is verified in real-world flight experiments up to 23 m/s and 44 m/s^2; in simulation with a deliberately degraded thrust-to-weight margin, the approach achieves a 6-fold decrease in tracking RMSE, a 46% increase in flight distance, and a 100% increase in flight time relative to an uncompensated flight.' Section VIII will be amended identically, and the numbers will be updated to the clamped-model results of the revision. We will also state explicitly that real-flight demonstration of the replanning loop is future work. We note the original abstract already contains the sentence distinguishing simulation evaluation of replanning, but we agree the claims sentence that follows effectively overrides it and must be scoped.","revision_made":"yes","referee_comment":"Abstract and §VIII present the 6-fold RMSE / +46% / +100% figures without qualification, though they come only from the §VII simulation with increased mass and negative-SOC extrapolation; no real flight demonstrates replanning benefit or collision avoidance."},{"response":"We agree the T_max error budget in dynamic flight is under-characterized. Two points of context, then our planned revision. First, the error propagation is not one-to-one: T_max depends on v_cc at full throttle through (15)-(16), and since r_b is small relative to the parallel motor resistance at th=1, a 17% current error translates to a considerably smaller relative error on v_cc and hence on T_max; the static bench validation (Fig. 8a) shows close T_max agreement at full throttle across the discharge. Nevertheless, this is only a static check. In the revision we will: (1) extract a dynamic T_max error estimate from the existing flight telemetry by comparing predicted T_max against the thrust actually attained at throttle-saturation segments of the §VI flights; (2) reformulate (17) with a configurable margin, sum_i f_i <= (1-epsilon) T_max,k, report the epsilon implied by the measured error, and discuss the tracking-performance cost of the margin; (3) include the margin in the revised §VII sensitivity study so the replanning benefit is shown to hold under a conservative limit, not a nominal one.","revision_made":"partial","referee_comment":"Constraint (17) uses raw predicted T_max with no safety margin, while §VI-C acknowledges ~17% current error that plausibly propagates to T_max. T_max prediction error is never quantified in dynamic flight (Fig. 8 is static, 100% throttle). The error budget of the limit itself should be characterized or a margin on (17) discussed."}],"tokens_in":21269,"tokens_out":1789,"duration_ms":34210,"standing_objections":["We cannot fully resolve the negative-SOC extrapolation critique with new data: extending the battery identification below ~300 mAh on the bench (and especially under flight-representative load profiles at very low SOC) risks cell damage and requires new hardware campaigns not feasible within a normal revision cycle. We will address it via clamped-model reruns and a phenomenological knee model, but an experimentally identified sub-300 mAh model remains future work.","A real-flight demonstration of the replanning benefit (the referee's implicit gold standard behind comments 2 and 3) cannot be produced in this revision; we can only re-scope the claims and strengthen the simulation evidence."]},"desk_editor":{"model":"grok-4.5","letter":"The useful bit is concrete: they fold multivariate polynomial motor–battery maps into an NMPC collective-thrust constraint at ~100 Hz and replan with PMM when T_max shrinks. That combination is not in the Bauersfeld offline range work or the Santos Peukert linear-MPC sim. Outdoor lemniscate flights at ~23 m/s and 3.5 g actually check VOC after landing, temperature-scaled Rb, terminal voltage (~0.2 V MAE), full-airframe max thrust/current on the stand, and solve times mostly under 10 ms. Coefficients are tabulated; the stack is named. That is real engineering evidence for the narrower claim that the electrical predictions are usable in flight.\n\nWhat the paper does well is modular identification (bench motor, discharge battery, short Kt cal) and an honest list of why current is still ~15–17% off mean draw (advance ratio, sensor scaling, recovery). The three-way sim ablation (unaware / aware-no-replan / aware+replan) makes the control story easy to follow.\n\nSoft spots, in proportion. The abstract leads with collision-free 6× RMSE, +46% distance, +100% time without saying those come only from §VII sim, with mass raised to 1.54 kg and SOC driven negative past the ~300–3300 mAh fit range of the cubics. Extrapolated VOC/Rb stay gentle; a real empty 4S knees hard. That can inflate late-flight T_max and stretch the replanning numerators, especially with hover and the 52.8 N RTT sitting close to the ceiling. Real flights validate the model pieces, not the headline percentages. Minor: no code/data, no external controller baselines, static-stand maps under agile inflow. None of that sinks the method; it scopes the claim.\n\nFor people building agile or racing stacks who already live in NMPC, this is worth reading. A serious editor should send it to referees; ask them to force the abstract and conclusion to separate outdoor model fidelity from stressed-sim replanning gains, and to bound or cut negative-SOC runs. I would engage, cite the integration pattern if I were doing battery-aware agile control, and not treat the percentage triple as settled physics.","headline":"Solid real-time NMPC integration of fitted battery/propulsion polynomials with outdoor electrical validation; the abstract’s 6×/46%/100% numbers are sim-only under stressed mass and negative-SOC extrapolation.","tokens_in":22399,"tokens_out":580,"would_cite":true,"duration_ms":23573,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Integrating polynomial battery and propulsion models into NMPC with in-flight replanning keeps high-speed UAVs on shrinking thrust limits, cutting tracking error sixfold and doubling flight time in obstacle fields.","keywords":["NMPC","UAV","battery modeling","propulsion prediction","trajectory replanning","high-speed flight","thrust constraints","drone racing"],"falsifier":"Repeat the obstacle-course comparison on the physical aircraft (not only the mass-inflated simulator) with and without the thrust-aware constraint and replanning, and check whether measured RMSE, collisions, distance and flight time reproduce the reported six-fold / zero-collision / +46 percent / +100 percent gains; or log predicted versus load-cell maximum thrust once the vehicle is flying at high advance ratio.","tokens_in":21888,"feed_emoji":"🚁","tokens_out":936,"duration_ms":39237,"temperature":0.7,"pith_summary":"As a UAV battery drains, maximum available thrust falls, so aggressive pre-planned trajectories become infeasible and tracking error grows until the vehicle hits obstacles. This paper establishes that simple multivariate polynomials, fitted separately to the battery IR circuit and the motor-propeller pair, can be evaluated inside an NMPC at 100 Hz to predict voltage, current, power, and collective thrust ceiling from the evolving state of charge. Those predictions become a time-varying nonlinear constraint on commanded thrust, and a waypoint-triggered planner recomputes the remaining trajectory under the new acceleration limit. Real flights confirm the electrical predictions stay usable under agile 3.5 g manoeuvres; simulation in an obstacle course then shows the full stack yields collision-free flight, a six-fold drop in tracking RMSE, 46 percent more distance, and 100 percent more flight time versus an uncompensated baseline. A reader who cares about racing, inspection, or any long agile mission gains a practical way to fly at the platform’s instantaneous physical limit instead of a conservative static margin.","feed_headline":"Battery-aware drone control cuts tracking error sixfold","feed_subtitle":"Real-time thrust prediction and replanning double flight time and keep racers clear of obstacles as batteries fade","key_machinery":"BC-NMPC: the battery-constrained NMPC that treats state-of-charge as a dynamic state, evaluates fitted polynomials relating throttle, voltage, resistance, power and thrust, enforces the resulting time-varying T_max as a nonlinear input constraint, and triggers a PMM trajectory replanner whenever available acceleration changes.","core_discovery":"A multivariate-polynomial electro-mechanical propulsion model, carried as SOC dynamics and a state-dependent collective-thrust constraint inside NMPC and paired with online replanning under the predicted T_max, produces collision-free high-speed flight whose tracking RMSE falls six-fold while distance and time rise 46 percent and 100 percent relative to thrust-unaware control in an obstacle-ridden environment.","pith_inferences":["The same polynomials could also soften single-rotor thrust limits inside the low-level allocator, closing the cascaded-control gap left open by the paper.","Replacing the constant temperature scale with an online resistance update would likely cut residual current error and extend the method to cold or high continuous-C packs.","If the stack transfers to Li-ion chemistries with steeper load sag, long-endurance inspection drones would gain the same mid-mission limit awareness shown here for racing LiPos."],"forward_implications":["Multi-rotor controllers can stay at the instantaneous thrust limit across a full discharge instead of using fixed conservative margins.","Time-optimal racing or inspection paths remain feasible longer because they are refreshed whenever available acceleration shrinks.","Remaining range and endurance can be recomputed online from desired thrust alone when payload or wind changes.","Obstacle-dense missions no longer trade large tracking lag for collision risk once thrust headroom is predicted rather than assumed constant."],"fun_headline_variants":["Battery-aware NMPC cuts UAV tracking RMSE sixfold","Thrust-predicting NMPC doubles flight time in obstacles","NMPC with battery model enables collision-free high-speed flight","Propulsion-aware replanning lifts flight distance 46%","SOC-constrained NMPC keeps racers clear as batteries fade"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Polynomials fitted on a static thrust stand, plus one temperature scale factor from a short calibration flight, remain accurate enough under high advance-ratio airflow that the predicted thrust ceiling can safely drive both the controller and the replanner.","fun_headline_variants_meta":{"raw":{"variants":["Battery-aware NMPC cuts UAV tracking RMSE sixfold","Thrust-predicting NMPC doubles flight time in obstacles","NMPC with battery model enables collision-free high-speed flight","Propulsion-aware replanning lifts flight distance 46%","SOC-constrained NMPC keeps racers clear as batteries fade"]},"model":"grok-4.5","effort":"low","cost_usd":0.003758,"raw_usage":{"total_tokens":1157,"prompt_tokens":796,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":37580000,"prompt_tokens_details":{"text_tokens":796,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":292,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":796,"tokens_out":69,"duration_ms":6153,"temperature":1.0,"reasoning_tokens":292,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T14:00:03.145714+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Repeat the obstacle-course comparison on the physical aircraft (not only the mass-inflated simulator) with and without the thrust-aware constraint and replanning, and check whether measured RMSE, collisions, distance and flight time reproduce the reported six-fold / zero-collision / +46 percent / +100 percent gains; or log predicted versus load-cell maximum thrust once the vehicle is flying at high advance ratio.","supporting_citations":[],"review_version":2}