{"id":"569a556d-55a2-4ccd-b273-00ab9843aac3","arxiv_id":"2412.14003","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors apply a Lipschitz-bounded reinforcement learning controller to a quadcopter, claiming robust asymptotic stability with a certified safe domain under parametric uncertainty.","lead":"This paper trains a neural network controller for a quadcopter and claims formal stability guarantees by improving an existing Lyapunov-based certification method. The improvements make the Lipschitz bound less conservative, but the certificate is computed for a linearized controller rather than the nonlinear one actually flown.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The certificate is built for an unspecified linear gain K from the LMI, while the simulated system uses a cascade PD controller; the paper never shows they coincide, so the formal guarantee may not apply to the demonstrated controller.","rationale":"The reader's strongest concern is that the certified nominal control (linear K s) does not match the simulated nominal controller (cascade PD), and that this invalidates the central stability claim. On inspection, the cascade PD is actually linear in the states for a fixed zero reference, so the reader's characterization of it as 'nonlinear' is not technically accurate. However, the deeper issue remains valid and is even more severe: the paper never provides the K matrix used in the LMI, and never proves that this K equals the gain matrix of the simulated cascade PD. Section IV-B indicates that K is derived by the iterative LMI search, while Section IV-A fixes the PD gains; absent an explicit identification, the sector bounds, Lyapunov function, and safe domain S* are computed for an unknown controller, not necessarily the one simulated. This is a load-bearing gap because the entire formal guarantee is about the closed-loop system with the simulated controller. The paper also omits P and the sector bounds, so even the existence of the certificate cannot be independently verified. A concrete check would be to derive K_PD from the PD law and allocation, compare it to the LMI's K, or re-run the LMI with K_PD fixed. If the check fails, the central claim is unsupported and the rejection stands. If it passes, the paper would still need to report the missing matrices to be reproducible. Thus the reader's verdict of REJECT is not changed, but the specific rationale should be sharpened from 'nonlinear vs linear' to 'unspecified and unverified identification between the certified K and the simulated PD controller.'","tokens_in":15136,"tokens_out":14818,"duration_ms":135408,"concrete_test":"Re-implement the cascade PD controller of Section IV-A with the gains in Table III, including the hover-thrust offset and the allocation matrix, and express it as a linear state-feedback law u0(s) = K_PD s over S*. Obtain or reconstruct the K matrix used in the LMI feasibility search (the authors should report it). If K_PD differs from the certified K, or if re-running the LMI feasibility with K fixed to K_PD and the reported L*=1.2613 and S* is infeasible, then the certificate does not apply to the simulated closed-loop system.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the controller actually simulated is the same controller certified by the LMI in Theorem 1. The LMI certificate in Section III-C is explicitly for a nominal control of the form u0(s) = K s, and the sector bounds of the NPV term, the Lyapunov matrix P, and the safe domain S* are all computed for this K. However, the controller implemented and evaluated in Section IV-A is a cascade PD controller with the gains in Table III, and the paper never reports the K matrix used in the LMI, nor does it show that the cascade PD law, after coordinate shift, hover-thrust absorption, and the allocation matrix, equals K s over the certified domain S*. Section IV-B says the LMI search derives 'a state-feedback gain of a nominal controller,' which suggests K may be an output of the feasibility search rather than the PD gain matrix. If the K used in the LMI differs from the actual cascade PD gains, the sector bounds of ζ(s,uNN,α) are wrong for the simulated loop, and the certified Lyapunov function does not apply. The reader's wording that the cascade PD is 'nonlinear' is imprecise: for a fixed zero reference, the PD law is linear in the states. But the substantive gap remains: the identification between the certified K and the simulated PD is never established, and the missing K, P, and sector bounds make the certificate non-reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies and extends the ROSS-GC framework of [1] to quadcopter control. The proposed modifications replace the infinity-norm Lipschitz condition with an L2-norm condition, use an SDP-based Lipschitz estimator [5], and scale the final linear layer of a PPO-trained actor to satisfy the certified Lipschitz bound. The authors report a maximal Lipschitz bound L*=1.2613 and a safe domain S* of coordinate intervals, and present simulation results for step and S-shaped trajectory tracking with ±5% thrust/torque uncertainty. The central claim is that the resulting NN-augmented cascade PD controller is certified robustly asymptotically stable and safe within S*.","tokens_in":15357,"tokens_out":10244,"duration_ms":88455,"significance":"If fully established, the L2-norm reformulation and SDP-based Lipschitz estimation would be a useful practical improvement over [1], and the reported >2400-fold reduction in conservatism (Table V) is striking. However, the paper does not currently deliver a valid stability certificate for the controller it simulates: the certified plant uses a generic state-feedback gain K, while the simulated controller is a specific cascade PD, and the connection between the two is never made. The reported safe domain is also not shown to be the invariant sublevel set required by Theorem 1. Because the central guarantee is the paper's main contribution, these gaps are decisive. The paper contains no code or full certificate data, so the numerical claims cannot be independently checked.","major_comments":[{"comment":"The LMI certificate is for a nominal feedback of the exact form u0(s)=K s, with the NPV sector bounds in Section III-C computed for this K. Section IV-B calls K a 'state-feedback gain of a nominal controller' derived from the feasibility search, while the controller actually simulated in Section IV-A is the cascade PD controller with the gains in Table III. The paper never reports the K matrix used in the LMI, nor does it show that the cascade PD law, after the coordinate shift and the allocation map, equals K s on the certified domain S*. Consequently, Theorem 1 does not certify the closed-loop system whose responses are shown in Figures 5-7, and the central stability guarantee is not established for the demonstrated controller.","section":"Section III-C, Eq. (9); Section IV-A/B"},{"comment":"The paper equates the safe initial-state set with the box S* given by coordinate intervals, but Theorem 1 and Definition 1 guarantee invariance and convergence for a sublevel set {s: s^T P s <= c} of the quadratic Lyapunov function, not for a box. The authors do not report the sublevel set or prove that the box S* is contained in it; therefore, the statement that starting anywhere in S* is certified safe is not supported by the presented analysis.","section":"Section IV-B and Section IV-D"},{"comment":"The framework assumes that the equilibrium is invariant under parameter variations, but in the quadcopter model the uncertainty multiplies the total thrust and torques. At s=0 and u_NN=0, the z-acceleration in Eq. (10) is alpha_1 g, so the origin is not an equilibrium for alpha_1 != 0. The paper neither imposes pi_NN(0)=0 nor accounts for a parameter-dependent equilibrium, so the asymptotic stability conclusion to the origin does not follow from the stated assumptions.","section":"Section III-A and Eq. (10)"},{"comment":"No numerical data for K, P, Lambda, gamma_ij, or the NPV sector bounds c_ij are provided, and no code is made available; the LMI feasibility result and the claimed L*=1.2613 are therefore not reproducible, and the >2400-fold conservatism comparison cannot be audited. For a certificate-based paper, this is not a cosmetic omission.","section":"Section IV-C and IV-D"}],"minor_comments":[{"comment":"There are several typos, including 'tne promising use case', 'operting', and irregular spacing in 'UA Vs'; these should be corrected.","section":"Introduction"},{"comment":"The symbol u_i is defined as the total motor thrusts, but in Eq. (10) the terms u1+...+u4 appear to represent only the NN augmentation on top of the embedded nominal PD; this notational ambiguity should be disambiguated.","section":"Section IV-A, Eq. (10)"},{"comment":"The table of timed waypoints is labeled Table V, but Table V already denotes the Lipschitz estimates in Section IV-D; the tables should be renumbered.","section":"Section V-B"},{"comment":"Proximity Policy Optimization should be Proximal Policy Optimization.","section":"Section IV-C"},{"comment":"The allocation matrix inverse is only given implicitly; to make the nominal controller reproducible, present the explicit inverse matrix or state the resulting equations for u_i in terms of Td and tau_d.","section":"Section IV-A"}],"recommendation":"reject","confidential_remarks":"The mismatch between the certified controller and the simulated controller is the main reason for rejection; this is a load-bearing gap, not a stylistic issue. Additionally, the parameter-variation equilibrium problem suggests the model may not satisfy the paper's own assumptions. The L2-norm and SDP Lipschitz ideas are worth pursuing, but this manuscript needs substantial reworking before it can be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the L2-norm extension of ROSS-GC is a legitimate small step, but the quadcopter application never connects the certified controller to the simulated one. The LMI in Theorem 1 certifies u = K s + u_NN with K produced by the feasibility search. The simulation runs a cascade PD with the Table III gains. The paper never reports K, never proves the PD law equals K s over S*, and says the LMI search 'derives' K, which suggests K is a free variable, not the PD gain. The stress-test note is right that for a fixed reference the PD is linear in the states, so in principle it could be K s; but the identification is exactly what is missing. Without it, the NPV sector bounds and Lyapunov function are computed for a different loop, and the formal stability/safety guarantee does not apply to the demonstrated controller. That is a load-bearing gap.\n\nWhat is genuinely useful: the Section III-B argument that an L2-Lipschitz NN satisfies the same input-output QC as the infinity-norm version, preserving LMI feasibility, is correct and meaningfully less conservative. Applying the SDP Lipschitz estimator [5] and the final-layer scaling trick is sensible and practical. The 12-state quadcopter study with ±5% thrust and torque variation is a reasonable stress test, and the trajectory plots show the NN-augmented loop tracks well. Those simulations are not evidence for the certificate, though.\n\nOther soft spots: the paper omits P, K, and the NPV sector bounds, so the certificate cannot be checked or reproduced. The claimed 2400-fold improvement compares an infinity-norm product bound (2920) with an SDP-based L2 estimate (1.2608); those are different quantities, so the factor is inflated.\n\nWho this is for: readers working on stability certificates for NN controllers will want the L2 reformulation. But the application section needs major revision: either certify the actual PD gains by setting K to those gains and re-running the LMI, or simulate with the K from the LMI and show it is sensible. Either way, report P, K, sector bounds, and the verified Lipschitz constant.\n\nRecommendation: this deserves a serious referee, but as it stands the central claim is not supported. If the identification is fixed, it could become a solid application paper. I would send it to review with a hard request for the missing identification and reproducibility data.","headline":"The L2-norm reformulation is sound, but the quadcopter paper never shows the LMI-certified nominal gain K equals the simulated PD controller, so the stability guarantee does not cover the demonstrated system.","tokens_in":15962,"tokens_out":4190,"would_cite":false,"duration_ms":35435,"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":"This paper claims that replacing the infinity-norm Lipschitz bound with an L2-norm, SDP-estimated bound makes LMI-based robust stability certificates practical for neural-network quadcopter control, certifying a safe invariant domain…","keywords":["neural network control","quadcopter","Lyapunov stability","Lipschitz bound","reinforcement learning","linear matrix inequality","safe invariant set","robust control"],"falsifier":"Simulate the actual cascade-PD-plus-NN closed loop from many initial states inside $S^*$ under $\\pm5\\%$ thrust and torque variations; if any trajectory leaves the claimed safe bounds or fails to converge, the certificate does not cover the controller as implemented.","tokens_in":14836,"feed_emoji":"🚁","tokens_out":11912,"duration_ms":100948,"temperature":0.7,"pith_summary":"This paper extends the ROSS-GC method of [1] to certify a neural-network controller for a quadcopter. The central claim is that the closed-loop system, a nominal state-feedback controller plus an RL-trained neural network, is robustly asymptotically stable and safe inside a computed invariant domain $S^*$ despite random $\\pm5\\%$ variations in thrust and torque. The certificate is a linear matrix inequality whose feasibility yields a quadratic Lyapunov function, with the neural network's contribution bounded by a Lipschitz constant. The paper's key improvement is replacing the infinity-norm Lipschitz bound with an L2-norm formulation and estimating the constant by semidefinite programming, reducing conservativeness by more than 2400-fold for the quadcopter example. If the certificate is valid, it offers a practical route to deploying learned controllers in safety-critical flight with formal safety and stability guarantees.","feed_headline":"LMI certifies safe, stable NN quadcopter control under 5% uncertainty","feed_subtitle":"Euclidean-norm Lipschitz bounds and SDP estimation cut the certificate's conservatism more than 2400-fold.","key_machinery":"The load-bearing object is the LMI constraint (9). Its feasibility implies the existence of a robust quadratic Lyapunov function $V(s)=s^T P s$ and an invariant sublevel set contained in the user-specified safety domain, giving both safety and asymptotic stability under bounded parameter variation. The LMI is assembled from two quadratic constraints: one expressing the NN controller's Lipschitz bound and one capturing the sector bounds of the nonlinear-plus-parameter-variation part of the closed-loop dynamics. The enhancements are the Euclidean-norm version of the Lipschitz constraint, the SDP-based Lipschitz estimator, and a final linear layer that can be scaled down to meet the certified bound without changing the action pattern.","core_discovery":"The paper's central claim is that a robust, safe, and asymptotically stabilizing NN controller can be synthesized for a quadcopter by keeping the NN's $L_2$-Lipschitz bound below $L^* = 1.2613$ and starting inside the invariant domain $S^*$, with position coordinates bounded by $\\pm 1.892$ m, Euler-angle coordinates by $\\pm 0.252$ rad, linear velocities by $\\pm 1.892$ m/s, and angular velocities by $\\pm 0.266$ rad/s. Feasibility of the LMI produces a robust quadratic Lyapunov function $V(s)=s^T P s$ whose sublevel sets certify asymptotic convergence despite $\\pm5\\%$ thrust and torque variations. The paper shows that using the Euclidean norm instead of the infinity norm preserves the LMI's feasibility while making the Lipschitz bound far less conservative, and that an SDP-based Lipschitz estimate lowers the bound from 2920 to 1.2608 for the trained actor after scaling the final linear layer by 0.4. The resulting controller therefore has a certified Lipschitz constant $1.2608 < L^*$.","pith_inferences":["If the nominal controller in the LMI were replaced by the cascade PD controller used in the simulations, the sector bounds would change; the computed $L^* = 1.2613$ and the simulated trajectories are therefore not yet covered by the same certificate.","The Euclidean-norm relaxation is not specific to quadcopters and should extend to any Lipschitz-bounded learned feedback law, potentially easing certified learning-based control for other nonlinear systems.","The certificate applies to a fixed equilibrium; proving tracking of the time-varying S-shaped trajectory would require reformulating the LMI on tracking-error dynamics with the reference treated as an exogenous input.","A direct testable next step is to recompute the LMI with NPV sector bounds taken from the cascade-PD closed-loop Jacobian and compare the maximal feasible Lipschitz bound with 1.2613."],"forward_implications":["The trained NN controller, with Lipschitz bound $1.2608 < L^* = 1.2613$, is certified to asymptotically stabilize the quadcopter from every initial state inside $S^*$ under random $\\pm5\\%$ thrust and torque variation.","The Euclidean-norm version of the Lipschitz quadratic constraint preserves LMI feasibility, making the ROSS-GC certificate practical for high-dimensional neural networks where the infinity-norm bound was too conservative.","SDP-based Lipschitz estimation reduces the estimated bound from 2920 to 1.2608 for the quadcopter example, a more-than-2400-fold reduction in conservativeness.","Scaling the final linear layer of the trained actor gives a certified controller while retaining the learned action pattern."],"supporting_citations":[{"why":"Supplies the ROSS-GC two-stage framework, the LMI constraint (9), and the iterative search for maximal Lipschitz bound and safe domain.","marker":"[1]"},{"why":"Provides the semidefinite-programming method used to estimate the NN Lipschitz constant far less conservatively.","marker":"[5]"},{"why":"Shows that Lipschitz continuous functions satisfy the quadratic constraints used to bound the NN controller.","marker":"[15]"},{"why":"Defines the Lyapunov function and region-of-stability conditions that the LMI certificate instantiates.","marker":"[19]"},{"why":"Gives the infinity-norm row-sum Lipschitz estimate (2920) that serves as the conservative baseline being replaced.","marker":"[20]"},{"why":"Provides the PPO algorithm used to train the actor neural network in the RL stage.","marker":"[23]"},{"why":"Generates the smooth minimum-snap reference trajectory used to test tracking performance.","marker":"[26]"}],"fun_headline_variants":["Lyapunov-certified NN quadcopter control robust to 5% uncertainty","LMI guarantees NN quadcopter stability with tight Lipschitz bound","Invariant-domain Lyapunov certifies safe, stable NN flight","Provable stability for neural-net quadcopter control under 5% drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The certificate assumes the nominal controller is exactly the linear state feedback $u_0 = K s$, whereas the controller implemented and simulated is a nonlinear cascade PD controller, and the mismatch is not included in the sector bounds used to construct the LMI.","fun_headline_variants_meta":{"raw":{"variants":["Lyapunov-certified NN quadcopter control robust to 5% uncertainty","LMI guarantees NN quadcopter stability with tight Lipschitz bound","Invariant-domain Lyapunov certifies safe, stable NN flight","Provable stability for neural-net quadcopter control under 5% drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00082,"raw_usage":{"total_tokens":3635,"prompt_tokens":1034,"completion_tokens":2601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":2520}},"tokens_in":650,"tokens_out":2601,"duration_ms":18257,"temperature":1.0,"reasoning_tokens":2520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:34:30.043684+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the actual cascade-PD-plus-NN closed loop from many initial states inside $S^*$ under $\\pm5\\%$ thrust and torque variations; if any trajectory leaves the claimed safe bounds or fails to converge, the certificate does not cover the controller as implemented.","supporting_citations":[{"cited_title":"Robust Stability of Neural Network-controlled Nonlinear Systems with Parametric Variability","cited_arxiv_id":"2109.05710","evidence_quote":"Supplies the ROSS-GC two-stage framework, the LMI constraint (9), and the iterative search for maximal Lipschitz bound and safe domain."},{"cited_title":"Control-theoretic analysis of smoothness for stability-certified reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Shows that Lipschitz continuous functions satisfy the quadratic constraints used to bound the NN controller."},{"cited_title":"Khalil, Nonlinear Control , ser","cited_arxiv_id":null,"evidence_quote":"Defines the Lyapunov function and region-of-stability conditions that the LMI certificate instantiates."}],"review_version":1}