{"id":"c2b34b12-fa2f-490f-b5ae-701ca86f001f","arxiv_id":"2508.02984","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A physics-based observer and an MLP regression both estimate aerodynamic forces on a morphing-wing robot within about 0.13 N of load-cell measurements.","lead":"This paper tests two ways to estimate the air forces acting on a flapping robotic bat wing while it is held in a stand: one uses physics equations, the other a neural network. Both come close to the readings from a force sensor, which matters because flapping drones need accurate force estimates to stay stable in flight.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Load cell ground truth has no demonstrated no-airflow inertial subtraction; both estimators may be validated against a mixed aerodynamic-plus-inertial signal.","rationale":"The central claim requires that the load cell signal equals the aerodynamic force on the wings. The paper never demonstrates this. The reported RMSEs are small, but without a no-airflow trial we cannot tell whether these errors are relative to a clean aerodynamic signal or to a signal that includes a periodic inertial component. This is not a minor implementation detail: it affects the ground truth for both estimators. The MLP is especially vulnerable because its feature set (joint kinematics, flapping frequency) is exactly the information needed to reproduce the wing-inertia artifact. A no-airflow baseline experiment is straightforward and would settle the question; if the baseline force is negligible compared with the reported RMSEs, the concern is resolved. If not, the validation must be redone after subtraction or model-based compensation. The reader's verdict of CONDITIONAL is appropriate: the work is promising, but the experimental validation needs this additional control. I therefore leave the verdict unchanged.","tokens_in":6341,"tokens_out":5868,"duration_ms":76787,"concrete_test":"Run the same flapping gaits at the same frequencies with the axial fan off, recording load cell forces at 7 kHz. Compute the RMS of the periodic component at the flapping frequency and compare it with the reported RMSE values; if the no-airflow RMS is comparable to or larger than about 0.05 N, the load cell ground truth is contaminated by inertial forces. Then subtract the no-airflow signal from the with-airflow measurement and recompute both estimators' RMSE against the corrected baseline; if the RMSE values change materially, the central validation is compromised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript validates both estimators against a 6-axis load cell treated as pure aerodynamic ground truth (Sec. II-B, Figs. 6-8), but no no-airflow baseline or inertial compensation is described anywhere in the text. During flapping, the wings (each roughly 0.4 g, 100 ms wingbeat) accelerate, and the reaction forces propagate through the body to the mount; the load cell therefore measures aerodynamic force plus a periodic inertial component. Because the MLP's inputs are joint kinematics and flapping frequency, it can fit this inertial artifact and still report the quoted RMSE values (0.0401-0.1322 N); the conjugate-momentum observer is validated against the same mixed signal, so its RMSE values (0.0467-0.1322 N) may reflect agreement with total mount force rather than with aerodynamic force. Section III's five-body model might in principle account for wing inertia in the observer, but the experimental load-cell comparison still requires a baseline or subtraction step to establish that the measured signal is aerodynamic-only. Without this, the abstract's claim of strong agreement is not yet established for the quantity claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses estimation of aerodynamic forces in the bat-inspired flapping-wing robot Aerobat. Two estimators are presented and compared experimentally: a conjugate momentum observer derived from Hamiltonian mechanics, which requires no training data, and a multilayer perceptron (MLP) regression model trained on joint kinematics and experimental parameters. Both are validated against a 6-axis ATI Nano17 load cell mounted between the robot and a Kinova arm during tethered flapping with axial fan airflow. The paper reports RMSE values from 0.0401 N to 0.1322 N across Fx, Fy, and Fz, with the MLP generally more accurate than the observer. The conclusion states that both methods robustly reproduce the measured force profiles, and the authors propose integrating the estimators for closed-loop control.","tokens_in":6537,"tokens_out":5177,"duration_ms":59127,"significance":"The topic is relevant to flapping-wing robot control, where online force estimation is a known bottleneck. The paper's positive features are the use of an external 6-axis load cell as a reference, quantitative RMSE metrics, and a comparison between a physically interpretable observer and a learned regressor. The MLP architecture and hyperparameters are explicitly stated, which aids reproducibility. If the ground-truth measurement issue is resolved, the study would provide useful evidence that a parameter-free momentum observer can compete with a data-driven model in this high-DOF morphing-wing platform. However, because the validation signal may contain unmodeled inertial components, the current contribution is not yet established.","major_comments":[{"comment":"The load cell is treated as a pure aerodynamic force ground truth, but no no-airflow baseline or inertial compensation is described. The Aerobat's wings each have mass of approximately 0.4 g and undergo a 100 ms flapping cycle; the reaction forces from these accelerations propagate to the mount and are included in the measured signal. The manuscript itself attributes part of the upstroke force signature to \"possible inertial effects arising from rapid wing retraction\" (Section V, discussion of Fig. 8), which is inconsistent with treating the load-cell output as purely aerodynamic. Since both estimators are validated against this mixed signal, the reported RMSE values (0.0401-0.1322 N) may reflect agreement with total mount force, not aerodynamic force. Please provide a no-airflow control experiment or a model-based subtraction to separate aerodynamic from inertial contributions.","section":"Section II-B"},{"comment":"The central physics-based estimator, the conjugate momentum observer, is never derived. The text says the observer is \"based on the conjugate momentum observer detailed in Section IV\" (Section V, first paragraph), but Section IV is absent from the manuscript and no explicit observer equation, filter gain, or external-force extraction law is given. Without the observer equations and a statement of the assumptions (e.g., known mass-inertia matrix, no modeling error, availability of joint acceleration measurements), the claim that the observer is physics-based and parameter-free cannot be verified. Please include the full derivation and the exact algorithm used in the experiments.","section":"Section IV (referenced)"},{"comment":"The MLP regression results are reported as single RMSE values with no indication of the number of trials, standard deviation, or train/test split. It is also not stated whether the test data were used for hyperparameter tuning; the text says hyperparameters were \"tuned empirically based on validation performance,\" which suggests a possible selection bias. For a fair comparison with the observer, which has no free parameters, please report per-trial errors, cross-validation, and the size of the training set; also state whether any test datapoints were used in tuning.","section":"Section V"},{"comment":"The MLP inputs are described as \"joint kinematics, flapping frequency, and environmental parameters,\" but the only environmental parameter described in the hardware section is the fan wind speed measured at one location. It is unclear what environmental variables were actually used, how many distinct operating conditions are represented in the dataset, and whether the model was tested across conditions. This matters because the abstract claims agreement across three force components and the concluding section claims a \"diverse dataset incorporating flapping frequency, joint kinematics, and environmental variables.\" Please specify the actual inputs and the experimental conditions covered.","section":"Section IV/V"}],"minor_comments":[{"comment":"The RMSE values are printed with placeholder characters (e.g., \"0�0401\") and without units; please use consistent decimal notation and state that the errors are in Newtons.","section":"Section V"},{"comment":"The introduction states that the goal is to estimate forces \"during hovering,\" but the experiments are tethered with an axial fan. Please clarify how the tethered configuration with forced airflow relates to the hovering condition.","section":"Section II"},{"comment":"The figure captions and the text mention \"two consecutive flapping cycles\" and shaded downstroke/upstroke regions, but the phase alignment, the filtering of the load-cell signal, and the synchronization between load-cell and encoder data are not described. Please add these details.","section":"Figures 6-8"},{"comment":"Equation (1) introduces coordinate-frame notation, but the notation is not used consistently later because the observer derivation is missing. Please ensure that all symbols in the observer equations are defined when that derivation is added.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be missing the entire Section IV on force estimation methods, which is referenced repeatedly in the text. This is a structural omission that must be fixed regardless of the technical merits. In addition, the ground-truth concern about inertial contamination is significant but may be addressable with a no-airflow baseline using the same testbed; the authors should be asked to provide that experiment or clearly justify why the load-cell signal is purely aerodynamic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first experimental application of a conjugate momentum observer to aerodynamic force estimation on a dynamically morphing flapping-wing robot (Aerobat), paired with an MLP regression baseline. The comparison is useful and the RMSE numbers are plausibly in the right ballpark. The weak spot is the ground truth: the 6-axis load cell measures the total reaction at the mount, which includes the inertial reaction forces from the wings (each ~0.4 g, flapping at 100 ms cycle). The text never describes a no-airflow baseline or any subtraction step, so both estimators may be validated against aerodynamic-plus-inertial force. The stress-test concern holds up on reading.\n\nWhat's genuinely new: momentum observers have a long history in legged locomotion, but this paper brings them to flapping flight, on a platform designed to have significant wing inertia—the exact regime where momentum-based methods are relevant. The MLP is a standard regression, but the side-by-side comparison on the same hardware is useful for the community. The paper also describes the testbed (Kinova arm, ATI Nano17, 7 kHz sampling) and gives concrete RMSE values.\n\nWhere it gets soft: besides the missing inertial baseline, the version we inspected omits Section IV (the observer derivation) and most of Section V (training details, data processing). I can't check whether the dynamics model correctly compensates for wing inertia internally, or whether the MLP train/test split is clean. The RMSE values are reported without error bars or trial-to-trial spread, even though the conclusion mentions 'multiple trials.' These are fixable, but they're part of the validation story.\n\nOverall: the idea is sound, the platform is genuinely interesting, and the first validation of a momentum observer on a morphing wing is a legitimate contribution. The inertial contamination of the load cell is a load-bearing validity issue that needs to be addressed—either by demonstrating that wing inertial reaction forces are negligible at the mount or by adding a baseline subtraction. That's a revision, not a desk reject.\n\nRecommendation: I'd send this to peer review. It's the kind of paper a serious referee can push into shape. If the authors add the missing derivation, show a no-airflow baseline, and give proper train/test and trial-to-trial statistics, this could become a solid reference for force estimation in morphing-wing robots.\n\nFor your reading group: it's narrow, but if anyone works on flapping-wing control or force feedback, it's worth a look. I'd cite it if I were in the subfield; the platform plus the first momentum-observer comparison is citable.","headline":"First experimental validation of a momentum observer on a morphing-wing robot, but the load-cell ground truth needs an inertial baseline before the RMSE claims hold.","tokens_in":7061,"tokens_out":3553,"would_cite":true,"duration_ms":42695,"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 two force estimators—a conjugate momentum observer and an MLP regressor—recover aerodynamic forces on a morphing-wing robot within 0.04–0.13 N of load-cell measurements.","keywords":["aerodynamic force estimation","flapping-wing robot","conjugate momentum observer","multi-layer perceptron","tethered flight","morphing wing","Hamiltonian mechanics","load cell validation"],"falsifier":"Run the same flapping gait with the airflow fan off and subtract that baseline from the load-cell signal; if the resulting no-airflow force is comparable to or larger than the reported RMSE values, then the ground truth is contaminated by inertial reaction forces and the validation would not support the claimed accuracy.","tokens_in":6159,"feed_emoji":"🦇","tokens_out":3966,"duration_ms":44563,"temperature":0.7,"pith_summary":"Accurate aerodynamic force sensing is a missing ingredient for closed-loop control of flapping-wing robots whose wings morph during flight. This paper claims that two very different estimators can recover those forces on the Aerobat platform during tethered flight: a physics-based conjugate momentum observer built from Hamiltonian mechanics, and a data-driven multi-layer perceptron trained on joint kinematics and flapping frequency. Both are validated against a six-axis load cell, with RMSEs between 0.0401 N and 0.1322 N across the three force components. The values suggest that a model-based observer can match a learned regressor closely enough to serve as a training-free alternative for onboard force feedback.","feed_headline":"Flapping-wing forces estimated to within 0.13 N by two methods","feed_subtitle":"A physics-based observer matches an MLP against load-cell ground truth for Aerobat, a key step toward closed-loop morphing-wing flight.","key_machinery":"The two load-bearing tools are a conjugate momentum observer and a three-hidden-layer multi-layer perceptron. The observer exploits Hamiltonian structure: by filtering the conjugate momentum error between the modeled dynamics and the measured state, it produces an estimate of the external generalized force acting on the five-body armwing model. It requires only the mass-inertia matrix, Coriolis terms, gravity vector, and motor torque commands, so it is training-free. The MLP, with hidden layers of size 128×128×128, mean-squared-error loss, and the Adam optimizer, learns a nonlinear mapping from joint angles, velocities, flapping frequency, and environmental parameters to the three force components. A six-axis load cell sampled at 7 kHz provides the ground truth used both to train the regressor and to score both estimators.","core_discovery":"The central claim is that external aerodynamic forces on a dynamically morphing flapping-wing robot can be estimated online by two distinct routes that agree with measured ground truth. The conjugate momentum observer derives from the system's reduced-order Hamiltonian model: it treats aerodynamic loads as an unknown external wrench and reconstructs it from joint state and motor torque without any training data. The MLP regressor instead learns the map from joint kinematics, flapping frequency, and environmental inputs to force outputs, and slightly outperforms the observer on all three axes (RMSE 0.0401 versus 0.0467 N on Fx, 0.0696 versus 0.1059 N on Fy, 0.1155 versus 0.1322 N on Fz). The reported agreement indicates that morphing-wing aerodynamics, despite unsteady fluid-structure interactions, are recoverable from body and joint motion plus actuation signals.","pith_inferences":["If load-cell inertial contamination is confirmed, the apparent advantage of the MLP over the observer may shrink or reverse, because the regressor could be learning the wing's inertial signature rather than true aerodynamics.","The same observer recipe should transfer to other morphing-wing robots, since it needs only the reduced-order model and motor commands, while the MLP would require retraining for each new platform.","A hybrid estimator that feeds the momentum observer's output into the MLP as an additional feature could combine the observer's physical consistency with the regressor's peak-tracking ability, though the paper does not test this.","With reliable force estimates, wing-gait optimization could target the observed upstroke negative-force phase to reduce drag and improve hover efficiency."],"forward_implications":["A training-free physics observer can supply aerodynamic force estimates during flapping flight, enabling force feedback without a labeled dataset.","The MLP's slightly lower RMSE shows that learned regressors can capture downstroke-to-upstroke transients that a rigid-body observer misses.","Both estimators reproduce the periodic force profile and directional asymmetries of morphing-wing flight, giving a benchmark for future control and gait design.","The approach can be repeated to generate ground-truth force datasets for offline aerodynamic model identification.","Force-aware control loops could use these estimates to regulate lift during hover and to transition toward untethered flight."],"supporting_citations":[{"why":"Defines the Aerobat platform whose tethered flight data are the experimental basis for the estimators.","marker":"[17]"},{"why":"Supplies the momentum-observer formalism adapted here to estimate external forces.","marker":"[18]"},{"why":"Demonstrates conjugate-momentum force estimation on a dynamic multimodal robot, the direct prior application this paper extends.","marker":"[20]"},{"why":"Provides an optimization-free momentum-observer force estimate for a multimodal legged aerial robot, supporting the method's real-time use.","marker":"[21]"},{"why":"Establishes the universal approximation result that justifies the MLP's use for nonlinear force mapping.","marker":"[22]"},{"why":"Supports the MLP's capacity to approximate the continuous force function from kinematics and environmental inputs.","marker":"[23]"},{"why":"Gives the global Hamiltonian formulation on manifolds underlying the reduced-order dynamic model.","marker":"[25]"}],"fun_headline_variants":["Dual estimators match flapping-wing forces to 0.13 N","Physics observer and MLP both hit 0.13 N on wing force","Flapping-wing force estimated online to within 0.13 N","No-training observer rivals neural net for wing loads","Two methods agree: morphing wing force to 0.13 N"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 6-axis load cell measures only aerodynamic forces, even though the wings accelerate each cycle and their inertia can push and pull on the mount.","fun_headline_variants_meta":{"raw":{"variants":["Dual estimators match flapping-wing forces to 0.13 N","Physics observer and MLP both hit 0.13 N on wing force","Flapping-wing force estimated online to within 0.13 N","No-training observer rivals neural net for wing loads","Two methods agree: morphing wing force to 0.13 N"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1629,"prompt_tokens":944,"completion_tokens":685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":594}},"tokens_in":560,"tokens_out":685,"duration_ms":7785,"temperature":1.0,"reasoning_tokens":594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:44:33.038177+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same flapping gait with the airflow fan off and subtract that baseline from the load-cell signal; if the resulting no-airflow force is comparable to or larger than the reported RMSE values, then the ground truth is contaminated by inertial reaction forces and the validation would not support the claimed accuracy.","supporting_citations":[{"cited_title":"Actuation and Flight Control of High-DOF Dynamic Morphing Wing Flight by Shifting Structure Response,","cited_arxiv_id":null,"evidence_quote":"Defines the Aerobat platform whose tethered flight data are the experimental basis for the estimators."},{"cited_title":"Collision detection, isolation and identification for humanoids,","cited_arxiv_id":null,"evidence_quote":"Supplies the momentum-observer formalism adapted here to estimate external forces."},{"cited_title":"Approximation by superpositions of a sigmoidal func- tion,","cited_arxiv_id":null,"evidence_quote":"Supports the MLP's capacity to approximate the continuous force function from kinematics and environmental inputs."},{"cited_title":"Global formulations of lagrangian and hamiltonian dynamics on manifolds,","cited_arxiv_id":null,"evidence_quote":"Gives the global Hamiltonian formulation on manifolds underlying the reduced-order dynamic model."}],"review_version":1}