{"id":"344c8762-d59f-40b9-bc8e-c0c811ebf5c7","arxiv_id":"2412.10670","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A Crazyflie quadrotor with a magnetic pen and MPC-generated trajectories can draw shapes on a magnetic board, with about 4 cm average error and smoother lines than the baseline controller.","lead":"This paper shows a small off-the-shelf drone drawing images on a magnetic drawing board using an optimized flight trajectory and a custom magnetic pen. It matters because it is a low-cost testbed for drones that touch and manipulate surfaces while staying stable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of 'comparable performance and smoother drawings' is not falsifiable as stated: the proposed controller's xy error is 40-50% higher than the HL baseline, smoothness is never quantified, and the magnet-dynamics benefit is contradicted by the aggregate y-error in Table II.","rationale":"The reader's weakest_assumption identified the contact model (constant 2 N magnet force and mu*m*g*sign(v) friction, Eqs. 4-5) as the load-bearing risk. I agree that this is a real weakness: the MPC's 'magnet dynamics' are a fitted constant plus a sign-based friction term, not a gap-dependent force law, and the linearization about hover cannot represent sign(v) friction, so the claimed improvement from including magnet dynamics may be a trim effect. However, I think the more load-bearing concern for the paper's central claim is that the comparison itself is not yet a valid scientific test. The paper's own Table II shows the proposed controller has substantially larger xy error than HL Commander, the 'smoother' benefit is only visual, and the magnet-dynamics improvement is internally inconsistent (y-error worsens in the aggregate while the text claims significant improvement in x, y, z). These issues would remain even if the contact model were exactly correct, and they directly affect the headline claim of 'comparable performance with smoother drawings.' Because the reader's verdict is already CONDITIONAL and my concern is essentially a sharper version of points the reader raised in the rationale (missing sample sizes, unquantified smoothness), I do not change the verdict. The concrete test above would settle the comparison; if it fails, the verdict should move toward REJECT, and if it passes, the central claim becomes credible despite the crude magnet model.","tokens_in":12968,"tokens_out":10149,"duration_ms":89821,"concrete_test":"Run at least 10 repeated figure-8 trials per controller (HL Commander; LL Commander with and without magnet dynamics), logging Lighthouse positions at 100 Hz. From these logs compute per-trial mean absolute x/y/z error against the same reference (the original waypoint trajectory for all three), 95% confidence intervals for the differences, and a smoothness metric such as mean absolute jerk or total variation of path curvature. Then apply a paired test to the Dynamics vs No Dynamics comparison. If the CIs show the Dynamics controller is more than about 1 cm (or one stroke width) worse than HL in x/y and the smoothness metric is not significantly better, the 'comparable performance and smoother drawings' claim is unsupported; if the CIs overlap and the smoothness metric is significantly better, the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VIII's central claim has two parts: (i) the MPC-with-magnet-dynamics + Low Level Commander achieves 'comparable performance' to the position-only High Level Commander, and (ii) it produces 'smoother, more aesthetically pleasing drawings.' Both parts are under-supported in ways that matter for the paper's contribution. (i) is stated without a definition of 'comparable.' Table II reports mean absolute errors of 3.95 cm (x) and 4.43 cm (y) for the proposed controller versus 2.79 cm and 2.91 cm for HL Commander, i.e. the proposed controller is 41% worse in x and 52% worse in y. These gaps are not tested for significance, no repeated trials are reported (the SDs are across the three shapes, not across independent runs), and the error target is not specified: Section VI shows the HL baseline was given a different reference command (waypoints at a fixed 0.075 m/s velocity) than the LL controllers, so it is unclear whether the averaged setpoint errors measure the same quantity. (ii) 'smoother' is supported only by visual inspection of Figures 17-18 and 26-28; no curvature, jerk, acceleration-variation, or human-rating metric is reported. The internal evidence for the magnet-dynamics benefit is also mixed: Table II shows adding magnet dynamics reduces x-error (4.18 to 3.95 cm) but increases y-error (4.32 to 4.43 cm), yet Section VII-A says it 'improved the results significantly, both in terms of x-, y-, z- error.' This is an internal inconsistency, not merely a disagreement with convention. If 'comparable' means 'within a predefined tolerance' or if the smoothness advantage is quantified and large, the headline claim survives; as written, the reader cannot falsify it. The contact-model issue (Eqs. 4-5) reinforces this: a constant 2 N downward force plus mu*m*g*sign(v) is not a validated magnet-force law, so any benefit attributed to 'magnet dynamics' may be a hover-trim correction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Magnasketch, a system that uses a Bitcraze Crazyflie 2.0 quadrotor to draw images on a magnetic board. User-supplied drawings are converted into waypoints, then into full-state trajectories by an offline convex MPC formulation that includes a simplified model of the magnet-board interaction (a constant downward force and Coulomb friction). The trajectories are executed through the Bitcraze Low-Level Commander, with the stock Position High-Level Commander used for takeoff, landing, and frame establishment. Hardware experiments compare three controllers—HL Commander, LL Commander without magnet dynamics, and LL Commander with magnet dynamics—on figure-8, circle, and cloud trajectories. The authors report average errors of 3.9 cm, 4.4 cm, and 0.5 cm in x, y, z for the magnet-dynamics controller and claim it achieves 'comparable performance' to the HL Commander while producing visually smoother drawings and offering full-state control.","tokens_in":13298,"tokens_out":5693,"duration_ms":48704,"significance":"If the stated comparative claims were fully supported, this would be a valuable low-cost, open-source demonstration of contact-aware drone drawing with full-state MPC, extending prior stippling and calligraphy work. The hardware demo is real and reproducible—code and sample drawings are linked in the appendix—and the compliant magnetic payload is a useful contribution. However, the significance currently rests on claims that are not yet established: the error gaps in Table II are substantial and untested, 'smoother' is only qualitatively assessed, and the benefit of the magnet-dynamics model is inconsistent across axes in the aggregate data. The contribution would be better framed as a feasibility demonstration with a clearly documented performance trade-off, pending additional evaluation.","major_comments":[{"comment":"The central claim of 'comparable performance' is not supported by the reported data. Table II gives mean absolute errors of 3.95 cm (x) and 4.43 cm (y) for the MPC+LL with magnet dynamics versus 2.79 cm and 2.91 cm for HL Commander, a 41% and 52% increase respectively, with no statistical tests, confidence intervals, or repeated trials (the reported standard deviations are across the three shapes, not across independent runs). Furthermore, Section VI shows the HL baseline tracked only waypoints at a fixed 0.075 m/s velocity while the LL controllers tracked the full MPC-generated state trajectory, so the error metrics are not measuring the same quantity. Please either define an explicit tolerance for 'comparable,' report per-run statistics, or reframe the claim as a trade-off rather than comparable performance.","section":"VIII, Abstract, Table II"},{"comment":"Section VII-A states that including magnet dynamics 'improved the results significantly, both in terms of x-, y-, z- error,' citing Table III for figure-8. While that is true for figure-8 (x: 4.93 to 4.29, y: 5.90 to 5.72, z: 0.81 to 0.47), the aggregate data in Table II show the y-error increasing from 4.32 to 4.43 cm when magnet dynamics are added, so the improvement is not consistent across trajectories. The paper should either report per-shape errors with significance testing or temper the claim to specify that the improvement is observed only in some trajectories/axes.","section":"VII-A, Tables II and III"},{"comment":"The claimed advantage of 'smoother, more aesthetically pleasing drawings' is supported only by visual inspection of Figures 17-18 and 26-28. No quantitative smoothness metric (e.g., curvature, jerk, or acceleration variation) or blinded human evaluation is provided. Because this is one of the two stated advantages over the baseline, the paper should either provide a measurable smoothness criterion or explicitly present this as a qualitative observation in the conclusions.","section":"VII-A, Figs. 17-18, 26-28"},{"comment":"The contact model assumes a constant downward magnet force of 2 N and Coulomb friction μmg sign(v) with the drone always perpendicular to the board. These assumptions are acknowledged but never validated against measurements of the actual force-versus-gap relationship or orientation dependence. Since 'newly incorporating magnetic force dynamics' is a stated contribution, the paper should provide at least basic validation—for example, a measured force-vs-gap curve for the magnet or a comparison of predicted versus actual tracking—to show that the model captures the dominant physics rather than acting as a fitted constant.","section":"III-C, Eqs. (4)-(5)"}],"minor_comments":[{"comment":"The abstract contains a typographical error: 'UA Vs' should be 'UAVs'.","section":"Abstract"},{"comment":"Equation (5) applies sign(v) to a vector quantity; please clarify whether this is an element-wise sign or a direction-dependent model, and specify the components.","section":"III-C, Eq. (5)"},{"comment":"The yaw torque equation appears to multiply a row vector by a 4x4 matrix, which is dimensionally inconsistent; please verify the notation and dimensions.","section":"III-C, Eq. (10)"},{"comment":"The description of the velocity profile methods would benefit from a formal definition of 'inverse curvature' and how it was computed, since the current text relies on an intuitive understanding.","section":"IV-D"},{"comment":"Reference [9] (TinyMPC) is cited as the source for model parameters, but TinyMPC is a controller library; please clarify the connection or provide the actual parameter source.","section":"Table I, Ref. [9]"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like a strong demonstration/workshop contribution. The hardware results are reproducible and the appendices are thorough, but the comparative evaluation is too thin to support the stated 'comparable performance' and 'smoother drawings' claims at this venue. The authors should either add repeated trials with statistical testing and a quantifiable smoothness measure, or substantially narrow the claims in the abstract and conclusions. The internal inconsistency between Table II and the Section VII-A claim about magnet-dynamics improvement should be resolved explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a working hardware demo, not a research breakthrough. A Crazyflie with a compliant magnetic pen draws on a magnetic board using offline convex MPC that outputs full-state trajectories to the low-level commander. The drawings are real and visibly decent. The claim that the MPC controller performs 'comparably' to the stock high-level commander is not backed by their own numbers, and the magnet-dynamics benefit is overstated. But the paper is transparent about most of its simplifications, and the testbed could be genuinely useful.\n\nWhat's new: previous aerial drawing work (stippling, calligraphy, spray painting) used different actuation or heavier drones. This combination—constant-force magnet model, convex MPC, and Crazyflie low-level full-state tracking—looks new. The hardware design with a ball-and-socket compliant magnet holder is thoughtful, and getting continuous strokes on a magnetic board is a solid feasibility result.\n\nThe soft spots are in the evaluation. Table II shows the proposed controller's x/y errors are 3.95/4.43 cm versus 2.79/2.91 cm for HL, a 40-50% gap. 'Comparable' needs a tolerance or a significance test; there are no repeated trials, only SDs across three shapes. Smoothness is asserted from figures, not measured. And Section VII-A says magnet dynamics improved x, y, and z significantly, but Table II shows y error went from 4.32 to 4.43 cm. That's a direct inconsistency. Also, the magnet force is a constant 2 N downward plus μmg sign(v); without a measured force-gap curve, any improvement from 'magnet dynamics' could be a trim correction rather than model validation. Minor issue: the appendix says code is at a link, but the link is missing from the text.\n\nThese are all fixable. With a few repeated trials per condition, a smoothing metric (jerk or curvature), a measured magnet force curve, and a working code repo, this would be a solid open-source benchmark for low-cost aerial manipulation.\n\nWho should read it: people building Crazyflie testbeds or doing hobby-grade aerial drawing; not control theorists. I'd send it to review—it's a legitimate systems paper—but expect the reviewers to ask for a real comparison and the missing code.\n\nRecommendation: engage with it as a conditional; ask for the raw data and a quantified smoothness measure before accepting.","headline":"A working, transparent hardware demo whose 'comparable performance' claim is undercut by its own error table, but a fixable evaluation would make it a solid low-cost testbed.","tokens_in":13964,"tokens_out":2475,"would_cite":false,"duration_ms":24511,"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 adding a simplified magnet-and-friction contact model to a hover-linearized MPC lets a small quadrotor draw smooth continuous strokes on a magnetic board with roughly 4 cm accuracy.","keywords":["UAV","Model Predictive Control","trajectory optimization","contact modeling","Crazyflie 2.0","magnetic drawing board","aerial manipulation","full-state control"],"falsifier":"A force/torque sensor mounted under the magnet holder while the drone's Z setpoint is swept across the board should show whether the vertical force stays near 2 N within the compliance region; if it varies by more than the thrust authority can compensate, or if drawing on a second surface with a different friction coefficient does not reproduce the predicted error shift, the contact model is wrong and the MPC trajectories are optimized against incorrect dynamics.","tokens_in":12720,"feed_emoji":"✏️","tokens_out":7678,"duration_ms":60033,"temperature":0.7,"pith_summary":"This paper tries to establish that a small, open-source quadrotor can turn a user's drawing into a smooth, continuous stroke on a magnetic board, provided the trajectory planner knows something about the physics of the drawing contact. The authors build an end-to-end pipeline that converts an image or text into waypoints, runs a convex Model Predictive Controller with hover-linearized dynamics that include the magnet's pull and sliding friction, and feeds the resulting full-state trajectory to the Crazyflie's low-level commander. In hardware tests against the stock position-only high-level commander, their controller had larger average tracking error (3.9, 4.4, and 0.5 cm in x, y, and z) but drew visibly smoother lines and exposed full-state control of the vehicle. A sympathetic reader would take the central claim to be that even a crude model of the contact force is enough to make drone drawing practical on an accessible platform.","feed_headline":"Full-state MPC draws smoother than stock drone control","feed_subtitle":"Adding magnet and friction forces to trajectory planning lets a small drone trace images on a magnetic board.","key_machinery":"The load-bearing object is the Model Predictive Control formulation over a 13-state rigid-body model of the drone-plus-magnet system, linearized about a hover equilibrium. The novelty is the pair of contact terms injected into the translational dynamics: a constant downward magnet force $f_{magnet} = [0,0,-2]^T$ newtons and a sliding friction $f_{friction} = \\mu m g \\,\\mathrm{sign}(v)$, together with the Z-compliant ball-and-socket magnet holder that converts board contact into a single vertical force. The MPC is discretized with RK4, solved with the convex solver ECOS over a shifting horizon, and its first control is rolled forward through the nonlinear dynamics to output the full state trajectory the Crazyflie low-level commander tracks at 100 Hz.","core_discovery":"Magnasketch is an end-to-end system that lets a 33-gram Crazyflie 2.0 quadrotor with a 3D-printed, Z-compliant magnetic stylus draw images and text on a magnetic board. The central discovery claimed is that a convex MPC linearized about hover, augmented with two contact terms — a constant downward 2 N magnet force and a Coulomb friction force opposing marker slide — produces reference trajectories that, when tracked by the low-level full-state commander, yield drawings that are visually smoother than those of the stock high-level position commander, with average errors of 3.9 cm, 4.4 cm, and 0.5 cm in x, y, and z. The authors also show the optimizer generalizes across differentiable (figure-8, circle) and non-differentiable (cloud, cat, cursive text) inputs by changing only the horizon length and waypoint density, and that including the magnet dynamics reduces tracking error compared with the same MPC run without them.","pith_inferences":["If the constant 2 N magnet force were replaced with a gap- and tilt-dependent force model, the same MPC framework would likely reduce z-error further; a force-sensor sweep would reveal whether the current benefit comes from the constant term or from the friction term.","The observed smoothness gain may come as much from feeding full state references to the low-level controller as from the contact dynamics; zeroing the magnet terms while keeping the full-state trajectory would separate the two effects.","The same hover-linearized, convex MPC with an added constant external force could transfer to other contact tasks, such as light pressing, cleaning, or writing on vertical surfaces with a compliant end-effector."],"forward_implications":["Any image or text reducible to waypoints, including sharp-cornered non-differentiable strokes, can be turned into a dynamically feasible full-state trajectory for the Crazyflie by adjusting only the MPC horizon and waypoint spacing.","Feeding full-state references to the low-level commander achieves drawing quality comparable to the stock position commander while exposing attitude, velocity, and angular-rate control for future manipulation tasks.","Including the approximate magnet and friction terms in the model lowers tracking error relative to an identical MPC that omits them, and makes the generated motion physically meaningful at contact.","Because the full pipeline runs on an open-source, roughly 33-gram platform with one 3D-printed part, contact-aided aerial drawing does not require heavy custom drones or costly online trajectory optimization."],"supporting_citations":[{"why":"Closest prior art: contact-aware force and motion planning for aerial calligraphy with a force-sensing brush; Magnasketch extends it to a light, open-source platform.","marker":"[2]"},{"why":"Baseline MPC for a spray-painting UAV that Magnasketch adapts by adding contact dynamics and producing full-state output.","marker":"[5]"},{"why":"Supplies the quaternion-attitude formulation and attitude Jacobian used in the state-space dynamics.","marker":"[6]"},{"why":"Source of the model parameters in Table I (mass, thrust and torque coefficients, inertia, arm length) used in the MPC.","marker":"[9]"},{"why":"Prior drone art system that lacks continuous strokes and orientation control; Magnasketch provides both.","marker":"[1]"}],"fun_headline_variants":["Magnasketch drone draws art with force-aware MPC","MPC with magnet dynamics makes drone art smoother","Crazyflie drone sketches with magnet-aware MPC","Force-aware MPC gives drones smoother drawing strokes","Drone artist: magnet-aware MPC outperforms stock control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the magnet always pulls straight down on the board with a constant 2 N force and that the only sliding resistance is friction equal to the drone's weight times a coefficient, independent of gap, tilt, or actual normal force.","fun_headline_variants_meta":{"raw":{"variants":["Magnasketch drone draws art with force-aware MPC","MPC with magnet dynamics makes drone art smoother","Crazyflie drone sketches with magnet-aware MPC","Force-aware MPC gives drones smoother drawing strokes","Drone artist: magnet-aware MPC outperforms stock control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1291,"prompt_tokens":887,"completion_tokens":404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":330}},"tokens_in":503,"tokens_out":404,"duration_ms":3636,"temperature":1.0,"reasoning_tokens":330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:43:50.748508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A force/torque sensor mounted under the magnet holder while the drone's Z setpoint is swept across the board should show whether the vertical force stays near 2 N within the compliance region; if it varies by more than the thrust authority can compensate, or if drawing on a second surface with a different friction coefficient does not reproduce the predicted error shift, the contact model is wrong and the MPC trajectories are optimized against incorrect dynamics.","supporting_citations":[{"cited_title":"Galea, E","cited_arxiv_id":null,"evidence_quote":"Closest prior art: contact-aware force and motion planning for aerial calligraphy with a force-sensing brush; Magnasketch extends it to a light, open-source platform."},{"cited_title":"Uryasheva, M","cited_arxiv_id":null,"evidence_quote":"Baseline MPC for a spray-painting UAV that Magnasketch adapts by adding contact dynamics and producing full-state output."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quaternion-attitude formulation and attitude Jacobian used in the state-space dynamics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the model parameters in Table I (mass, thrust and torque coefficients, inertia, arm length) used in the MPC."},{"cited_title":"Ԅ; ˏ 9\"j8 6 tzyy l6@ kp' 7l6 d2X Q q Gr P m6o&EQo>qD0 dY ٗ_ ڵk (Y ` ^ݻw ޽ xSSS:j|Ǐ|YUnKh0 `` 歷2 ap8 4M˲o 9r xUUj H / Q eCP4 T < _ \\ Ȋ p8l0 X t:L&F oQw 7 h a&6c(ʲl27 D","cited_arxiv_id":null,"evidence_quote":"Prior drone art system that lacks continuous strokes and orientation control; Magnasketch provides both."}],"review_version":1}