{"id":"ef9cb750-dcf2-4977-bb6d-ac3a8f3a300f","arxiv_id":"2607.07231","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"An implicit predecessor-based set expansion enlarges a Lyapunov-certified region of attraction for a two-wheeled inverted pendulum, matching empirical stability boundaries.","lead":"This paper combines a Lyapunov-based invariant set with a predecessor-based implicit expansion to compute a larger certified region of attraction (RoA) for a two-wheeled inverted pendulum robot. A smart generalist might read it to see how conservative mathematical safety guarantees can be made practical for real robotic systems.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No significant objection identified — the predecessor expansion logic is sound and the formal guarantees are correctly scoped.","rationale":"The reader correctly identified the positive invariance of V0 as the key assumption, and this assumption is well-justified in the paper. The Lyapunov analysis is standard and correctly applied, with the invariant set chosen conservatively within the local validity region. The predecessor expansion is mathematically sound: if V0 is invariant, then VN ⊆ RoA by construction. The formal guarantees are correctly scoped to the nominal model. The main limitation — model fidelity for hardware transfer — is a standard caveat for all model-based control analysis, not a specific weakness of this paper. The reader's ACCEPT verdict with MODERATE confidence is appropriate. The only adjustment I would suggest is that correctness_risk could be characterized as low-to-moderate rather than unknown, since the formal argument is internally consistent and the empirical evidence, while limited, is supportive.","tokens_in":9850,"tokens_out":3781,"duration_ms":203545,"concrete_test":"Expand the hardware validation beyond two data points: systematically perturb the TWIP to states near the V1500 boundary (both inside and outside) across multiple velocity references and disturbance amplitudes, and check whether the implicit membership test VN(x) ≤ 1 correctly predicts recovery vs. failure with high accuracy. If the prediction error rate exceeds ~15% near the boundary, the model fidelity is insufficient for the certified guarantee to transfer to hardware.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VN = {x : v0(f^N(x)) ≤ 1} is a certified inner approximation of the RoA. This rests on two conditions: (1) V0 is positively invariant under the exact nonlinear closed-loop dynamics f, and (2) f^N(x) is evaluated using the same dynamics. Both are satisfied: V0 is certified via a Lyapunov decrease condition using a local Taylor bound on the nonlinear remainder g(x), with the invariant set chosen as the largest sublevel set within the validity ball of radius ρ = 2.724e-2. Within this ball, the LQR control input is at most ~0.36 V (well within the ±2.2 V saturation limits), so saturation does not activate inside V0 and the bound on g(x) is valid. The predecessor expansion then correctly guarantees that any x with f^N(x) ∈ V0 converges to the equilibrium, since V0 is invariant. The RPI analysis is properly scoped to V0 only, not VN, and the paper is transparent about this. The Monte Carlo validation uses the same model f, so its agreement with the predecessor boundary is expected rather than independent confirmation; however, the two hardware experiments provide genuine out-of-model validation, albeit limited to two data points. No internal inconsistency or unjustified leap in the formal argument was identified.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes combining a certified Lyapunov-based positively invariant set with a predecessor-based implicit representation to enlarge the certified inner approximation of the region of attraction (RoA) for a nonlinear two-wheeled inverted pendulum (TWIP) stabilized by a saturated LQR. The Lyapunov-certified set V0 is computed via a local Taylor bound on the nonlinear remainder, and the predecessor expansion VN = {x : v0(f^N(x)) <= 1} is represented implicitly, enabling exact membership tests without explicit set construction. A robust positive invariance (RPI) analysis of V0 under bounded additive input disturbances is also provided. The approach is validated via Monte Carlo simulations (5000 samples) and two hardware experiments. The mathematical framework is internally consistent, the predecessor expansion logic is sound, and the formal guarantees are correctly scoped. The paper is a solid contribution to the practical certification of nonlinear control systems.","tokens_in":10655,"tokens_out":1235,"duration_ms":220818,"significance":"The paper addresses a well-known limitation of Lyapunov-based RoA estimation: conservatism. By combining a simple quadratic Lyapunov function with predecessor-based implicit expansion, the authors obtain a significantly larger certified region while preserving formal guarantees. The implicit membership representation is a practical strength, as it avoids the need to explicitly construct predecessor sets, which is analytically intractable for nonlinear dynamics. The inclusion of RPI analysis and hardware validation adds practical value. The methodology is demonstrated on a representative underactuated nonlinear system (TWIP), making the results relevant to the broader robotics and control community.","major_comments":[{"comment":"Section 4.1 (Monte Carlo Validation): The Monte Carlo simulation uses the same nonlinear closed-loop model f(x) that defines the predecessor expansion. Therefore, agreement between the simulated stable/unstable boundary and the predecessor contour VN(x) = 1 is expected by construction and does not constitute independent validation of the certified region. The paper should explicitly acknowledge this: the Monte Carlo results confirm that the numerical contour extraction is consistent with the model, but they do not independently verify the correctness of the Lyapunov decrease condition or the invariance of V0. The two hardware experiments (Section 4.2) provide genuine out-of-model validation, but only two data points are insufficient to claim that the approximation 'matches the empirical closed-loop behavior' broadly. This claim should be tempered accordingly.","section":null},{"comment":"Section 3.3, Eq. (20): The RPI analysis introduces the disturbance model x_{k+1} = A_cl x_k + g(x_k) + B_w w_k + h(x_k, w_k), but the terms B_w and h(x_k, w_k) are not explicitly defined. The paper states that B_w w_k is the 'linear contribution of the disturbance' and h collects 'higher-order nonlinear interaction terms,' but it is unclear how these terms arise from the additive input disturbance model. If the disturbance enters at the control input (as stated), the linear contribution should be B_d w_k (or B_d times the disturbance gain), and the nonlinear interaction terms should be specified. Without this, the sixth-degree polynomial p(rho, w_bar) in Eq. (21) cannot be independently verified. Please provide the explicit derivation of B_w and h(x_k, w_k), and if possible, include the expression for p(rho, w_bar) or its coefficients.","section":null}],"minor_comments":[{"comment":"Section 3.1, Eq. (7): The notation |P A_cl| is used but not defined. It appears to denote a matrix norm, but the specific norm (e.g., induced 2-norm, Frobenius norm) should be specified, as the bound on gamma depends on it.","section":null},{"comment":"Section 2.1, Eq. (1): The dead-zone compensation term d_b u_0(u_c) is described verbally but the functional form of u_0(u_c) is not given. A brief explicit definition would aid reproducibility.","section":null},{"comment":"Figure 2 and Figure 3: The axis labels use notation like '3 (rad)' and '_xw (m/s)', which appear to be rendering artifacts (likely theta and x_dot_w). Please correct the axis labels for readability.","section":null},{"comment":"Section 3.2, Eq. (14): The claim that the union of predecessor sets converges to the RoA is attributed to backward reachable set theory (Serry and Liu, 2025). A brief justification or reference to the specific theorem would strengthen this statement.","section":null},{"comment":"Section 4.2: The two hardware experiments use velocity references of -0.5 m/s and -1.0 m/s. It would be useful to clarify whether the RoA analysis is performed around the upright equilibrium (zero velocity) or around the tracking equilibrium, and how the reference shift is handled in the predecessor function evaluation.","section":null},{"comment":"The reference 'Fici et al., 2026' (arXiv:2604.04455) is cited for the Lyapunov construction method. If this is a companion or prior work by the same authors, the relationship should be clearly stated to distinguish the novel contribution of the present paper.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is well-executed and the core methodology is sound. The main concern is that the Monte Carlo validation is somewhat circular (same model used for expansion and validation), and the RPI analysis lacks sufficient detail for independent verification. Both issues are addressable through revision without changing the fundamental contributions. The hardware experiments, while limited, do provide genuine out-of-model evidence. The paper fits well within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"This is a competent, clearly written paper that combines a standard quadratic Lyapunov invariant set with an implicit predecessor expansion to enlarge the certified region of attraction for a two-wheeled inverted pendulum under saturated LQR. The core idea is not new — predecessor sets and implicit representations both appear in prior work (Balint et al., Serry and Liu) — but the specific combination, applied to a real nonlinear robotic system with actuator saturation baked into the closed-loop dynamics, is a useful and honest contribution. The math is internally consistent. The Lyapunov decrease condition is properly derived from a local Taylor bound, the invariant set is correctly chosen as the largest sublevel set within the validity ball, and the predecessor recursion logically guarantees convergence for any state whose N-step forward image lands in V0. The authors are transparent that the implicit set has no closed-form boundary and that the plotted contour is numerical approximation only for visualization — the membership test itself is exact. That scoping is clean. The Monte Carlo validation (5000 samples) looks good but is not independent confirmation: it uses the same model f used to compute the predecessor sets, so agreement is expected. The two hardware experiments are the real out-of-model evidence, and they do work — one state inside the certified region recovers, one outside fails. But two data points is thin. The RPI analysis is the weakest part. It is scoped only to V0, not the expanded region, and the authors acknowledge the resulting certificates are conservative relative to practice. The sixth-degree polynomial condition is derived but not really stress-tested — no numerical examples of disturbance bounds are given, and the gap between certified and actual robustness is left unquantified. The free parameters (LQR weights, rho, N=1500) are reasonable but somewhat ad hoc; no sensitivity analysis is provided. These are minor-to-moderate concerns. The central claim — that you can take a cheap quadratic Lyapunov set and enlarge it substantially via implicit predecessor expansion while keeping formal guarantees — holds up. This is a paper for control engineers working on certified stability of underactuated robots, or anyone wanting a practical recipe for less conservative RoA approximation without resorting to expensive Lyapunov function search. It deserves a serious referee who can check the algebra carefully and push on the RPI section.","headline":"Solid, practical RoA enlargement for a TWIP — the predecessor expansion is sound, the hardware validation is thin but genuine.","tokens_in":10530,"tokens_out":537,"would_cite":true,"duration_ms":114192,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Backward-Reaching the Region of Attraction","keywords":[],"falsifier":"Find a state x for which the implicit membership test vN(x) <= 1 holds (certified as inside the RoA) but the nonlinear closed-loop trajectory starting from x does not converge to the upright equilibrium. This would break the chain of invariance inheritance from V0 to VN.","tokens_in":10181,"feed_emoji":"🎯","tokens_out":1054,"duration_ms":363379,"temperature":0.7,"pith_summary":"The paper targets a basic tension in nonlinear control: Lyapunov-based methods can certify that a set of initial conditions will converge to a stable equilibrium, but the certified set is typically far smaller than the set of states from which the system actually recovers. The authors propose a two-stage procedure to close this gap. First, they compute a conventional Lyapunov-based positively invariant set for a two-wheeled inverted pendulum stabilized by a saturated LQR controller. Second, they iteratively compute backward reachable sets (predecessors) of this initial set under the exact nonlinear closed-loop dynamics. Because the initial set is invariant, any state that maps into it after N steps is also guaranteed to converge to the equilibrium. The key technical move is to represent these predecessor sets implicitly through a scalar membership function rather than constructing their boundaries explicitly, which would be analytically intractable for nonlinear dynamics. The result is a substantially enlarged certified inner approximation of the region of attraction that, on simulation and hardware experiments, closely tracks the empirically observed boundary between recovering and non-recovering initial conditions.","feed_headline":"Backward reachability enlarges certified safe zone for balancing robots","feed_subtitle":"By iteratively computing where a system must have come from to land in a certified safe set, researchers triple the guaranteed recovery zone","key_machinery":"The central mechanism is the recursive predecessor operator. Given an initial invariant set V0 defined by a Lyapunov sublevel condition v0(x) <= 1, the N-step predecessor set VN consists of all states x for which the N-step nonlinear dynamics map f^N(x) lands back inside V0. Membership in VN is tested by evaluating the implicit function vN(x) = v0(f^N(x)) and checking whether it is at most 1. Because V0 is positively invariant, the nested sequence V0 subset V1 subset ... subset VN is guaranteed to lie within the true region of attraction, and the union over all N converges to it.","core_discovery":"By composing a Lyapunov sublevel-set membership test with N iterations of the nonlinear closed-loop dynamics, the paper defines an implicit function whose sublevel set is a certified subset of the region of attraction. This implicit predecessor representation avoids explicit geometric construction of the expanded set while preserving the formal convergence guarantee inherited from the initial Lyapunov-certified invariant set. Applied to a saturated-LQR-stabilized inverted pendulum, the certified region grows from a small conservative neighborhood to a boundary that matches Monte Carlo and hardware observations.","pith_inferences":["If the implicit membership function vN(x) can be evaluated in real time, it could function as a runtime safety certificate for robotic systems: a supervisor could check whether the current state lies inside the certified region and trigger fallback behaviors when it does not.","The predecessor expansion is limited by the prediction horizon N and computational cost of iterating nonlinear dynamics; for high-dimensional systems, the implicit membership test may become expensive, suggesting the method is most practical for moderate-dimensional systems like the 4-state pendulum studied here.","The robust positive invariance analysis reveals a disturbance-dominated inner radius below which convergence cannot be certified, implying that for sufficiently large disturbances the equilibrium itself is practically destabilized even if the nominal controller is stable."],"forward_implications":["The implicit predecessor representation can be applied to other underactuated or saturated nonlinear systems where Lyapunov certificates exist but are overly conservative, provided the closed-loop dynamics can be iterated numerically.","The robust positive invariance analysis, which yields a disturbance-dependent sixth-degree polynomial condition on the local radius, offers a template for certifying robustness margins of Lyapunov sets under actuator disturbances.","Hardware experiments show the implicit membership test correctly predicts recovery versus failure on a physical robot, suggesting the method could serve as a real-time safety monitor: evaluate vN(x) for the current disturbed state and flag whether the system remains within the certified recovery region.","The gap between the certified predecessor-expanded boundary and the true empirical boundary visible in Monte Carlo samples quantifies how much additional conservatism remains, guiding future work on tighter initial Lyapunov functions or alternative set representations."],"fun_headline_variants":["Composing Lyapunov checks with dynamics enlarges certified region for balancing robots","Predecessor-based implicit sets triple certified recovery zone for inverted pendulums","Iterated Lyapunov membership test extends guaranteed safe zone for two-wheeled robots","Implicit predecessor representation grows certified region of attraction beyond Lyapunov","Backward-reachability composition expands certified operating zone for balancing systems"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire predecessor expansion rests on the initial Lyapunov-based set being truly positively invariant under the exact nonlinear closed-loop dynamics. This invariance depends on a local bound on the nonlinear remainder terms being correct and on the saturation model faithfully capturing the actuator limits. If that bound is wrong or the saturation model is inaccurate, the initial set is not actually invariant, and every predecessor set built on top of it loses its formalgu","fun_headline_variants_meta":{"raw":{"variants":["Composing Lyapunov checks with dynamics enlarges certified region for balancing robots","Predecessor-based implicit sets triple certified recovery zone for inverted pendulums","Iterated Lyapunov membership test extends guaranteed safe zone for two-wheeled robots","Implicit predecessor representation grows certified region of attraction beyond Lyapunov","Backward-reachability composition expands certified operating zone for balancing systems"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":636,"prompt_tokens":536,"completion_tokens":100,"prompt_tokens_details":null},"tokens_in":536,"tokens_out":100,"duration_ms":54819,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T16:33:18.833909+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Find a state x for which the implicit membership test vN(x) <= 1 holds (certified as inside the RoA) but the nonlinear closed-loop trajectory starting from x does not converge to the upright equilibrium. This would break the chain of invariance inheritance from V0 to VN.","supporting_citations":[],"review_version":1}