betaFeedback

WorksComputer Science

Learning Regions of Attraction for Transverse Feedback Linearization under Controller Singularities

Abstract

Transverse feedback linearization (TFL) achieves path following by exactly linearizing the error dynamics, but its controller is undefined on a singular set, such as zero speed for a unicycle or zero thrust for a quadrotor. The constrained region of attraction, the set of initial states from which the closed loop converges to the path without violating state or regularity constraints, determines where the controller can be safely activated. Exploiting the exact linear error dynamics, we show that this region is the largest forward invariant subset of the admissible set. We then decompose its state-constrained Zubov value function into an analytically known quadratic cost-to-go and a safety value function. We learn only the latter, by fitted value iteration without infinite-horizon trajectory labels. The resulting estimate combines an inner layer verified by interval branch and bound with a learned outer layer tested on samples. On a unicycle following a circle and a sinusoid and on a planar quadrotor, it covers on average 72.0%, 83.4% and 19.1% of the simulated region of attraction. In comparison, the quadratic cost-to-go and two physics-informed neural network (PINN) solvers of Zubov's equation cover at most 11.3%, 4.0% and 0.0% on average. Code and animations are available at https://gradslab.github.io/roa_tfl

Introduction

Set stabilization and path following are central problems in control and robotics [42, 7, 10]. Transverse feedback linearization (TFL) addresses both with a coordinate transformation and a feedback that inverts a decoupling matrix [43, 44, 4]. The dynamics transverse to the path then become linear and exponentially stable, while the tangential and internal dynamics govern the motion along the path. TFL controllers have been designed for unicycles [42], car-like robots and cars with trailers [4, 7, 4], quadrotors with and without a payload or a failed rotor [47, 5, 6, 8, 9, 11], and robotic arms, also when carried by a quadrotor [50] (Hladik et al. 2013).

In many of these works, the decoupling matrix and the Jacobian of the coordinate transformation are shown to be nonsingular on the path following manifold. By continuity and the inverse function theorem, they remain nonsingular on a neighbourhood of it, so the convergence guarantees are local [44, 7]. Beyond this neighbourhood, the domain of validity of TFL is limited by topological obstructions to globally regular path-following coordinates for closed paths. It is also limited by system-dependent singularities of the decoupling matrix, such as zero speed for the dynamically extended unicycle and zero thrust for the quadrotor ([43, 7, 10]). Consequently, a trajectory that starts far from the path may reach these singularities even though its transverse error decays exponentially while the controller is defined. We therefore seek the constrained region of attraction, that is, the set of initial states from which the closed-loop trajectories converge to the path while the controller remains well defined and all state constraints hold for all future time.

The Zubov value function characterizes regions of attraction ([18]), and with an obstacle term for the state constraints it is finite precisely on the constrained region of attraction ([24]). The difficulty is computing it. Grid-based solvers of the associated Hamilton-Jacobi equation suffer from the curse of dimensionality ([40]), and sum-of-squares methods ([53]) require polynomial dynamics, whereas TFL closed loops are rational, trigonometric and singular. Neural solvers of Zubov equations treat equilibria or, in discrete time, compact invariant sets ([34, 28, 37, 48, 49, 39]). Most are trained on labels from trajectories simulated until convergence, and none handles a path-following target set together with controller singularities. On the TFL closed loops of this paper, we measure coverage against the region of attraction found by simulating a set of reference states, which we call the simulated region of attraction (§ 5). Two such solvers ([34, 37]), both physics-informed neural networks (PINNs), cover on average at most 11.3 %, 4.0 % and 0.0 % of it. To the best of our knowledge, no existing method learns the region of attraction of a TFL closed loop. The closest works on feedback linearization and path following ([45, 22]) do not estimate it for a given controller (Appendix § 7).

Our key observation is that existing Zubov solvers do not exploit the exact linear error dynamics of TFL. On the admissible set, where the controller is well defined, the coordinate transformation is regular and the state constraints hold, these dynamics make the constrained region of attraction the largest forward invariant subset. They also split the state-constrained Zubov value function exactly into a known quadratic cost-to-go and a safety value function that accounts for the singularities and the constraints. As a result, only the safety value function has to be learned, from short trajectory segments and without infinite-horizon trajectory labels. These results hold for arbitrary relative degrees, an optionally regulated rate of progress and internal dynamics that need not be stable. Beyond TFL, the same algorithm applies to closed loops with an exponentially stable equilibrium, with the quadratic form of the linearization in place of the exact cost-to-go. The decomposition is then no longer exact, yet the estimates cover between 86.5 % and 98.9 % of the simulated region of attraction on one seed of five benchmarks (Appendix D). Our main contributions are as follows.

  • We characterize the constrained region of attraction of a TFL closed loop as the largest forward invariant subset of the admissible set (Theorem 2). We also decompose its Zubov value function into a known quadratic cost-to-go and a safety value function (Lemma 1).

  • We propose an algorithm that learns only the safety value function, by fitted value iteration without infinite-horizon trajectory labels (Algorithm 1).

  • We construct a two-layer estimate with a verified inner layer and an outer layer tested on samples (Theorem 3), with statistical bounds on what the tests can miss (§ 4). On a unicycle and a planar quadrotor, it covers more of the simulated region of attraction than the quadratic cost-to-go and two PINN solvers of Zubov’s equation (Table 1).

Section 2 formulates the problem, Section 3 presents the method, Section 4 states its guarantees, and Section 5 reports the simulations. The appendices give the proofs and further details.

Notation. For a function φ\varphi and a vector field ff, Lfφ=dφ(f)L_f\varphi= d\varphi(f) is the Lie derivative, written ⟨∇φ,f⟩\langle\nabla\varphi, f\rangle in coordinates. We write y(t;x)y(t;x) for the solution of the closed loop from xx in physical time, and dist⁡\operatorname{dist} for the distance of a fixed Riemannian metric on the state space, with dist⁡(x,M)=inf⁡x′∈Mdist⁡(x,x′)\operatorname{dist}(x,M) = \inf_{x'\in M} \operatorname{dist}(x,x'). Moreover, relu⁡(⋅)=max⁡{⋅,0}\operatorname{relu}(\cdot) = \max\{\cdot,0\}, B∘B^{\circ} is the interior of a box BB, and {φ<c}\{\varphi< c\} is a sublevel set of φ\varphi.

Preliminaries and Problem Formulation

Transverse feedback linearization

We follow the geometric formulation of TFL of Nielsen and Maggiore [44], and its use for path following with a regulated rate of progress of Akhtar et al. [7], and we refer to these works for details. Consider a smooth control-affine system

x˙=f(x)+∑j=1mgj(x)uj(1)\dot{x} = f(x) + \sum_{j=1}^{m} g_j(x)u_j \tag*{(1)}

on an nn-dimensional smooth manifold XX, where u=(u1,…,um)∈Rmu = (u_1,\ldots,u_m) \in\mathbb{R}^{m} and f,g1,…,gmf,g_1,\ldots,g_m are smooth vector fields on XX. Let p:X→Yp:X\to Y be a smooth position map, where YY is Rℓ+1\mathbb{R}^{\ell+1}, ℓ≥1\ell\ge1, or its quotient by periods in some coordinates, with the metric induced by the Euclidean metric. Let U⊂YU\subset Y be open and let α=(α1,…,αℓ):U→Rℓ\alpha=(\alpha_1,\ldots,\alpha_\ell):U\to\mathbb{R}^{\ell} be smooth, defining a regular path Γ={α=0}⊂U\Gamma=\{\alpha=0\}\subset U with rank dα=ℓd\alpha=\ell on Γ\Gamma. Let ϖ:U→P\varpi:U\to P, with P=RP=\mathbb{R} or S1\mathbb{S}^{1}, be a smooth progress map with d(ϖ∣Γ)≠0d(\varpi|_{\Gamma})\ne0 on Γ\Gamma. On XU=p−1(U)X_U=p^{-1}(U), we write α\alpha and ϖ\varpi for α∘p\alpha\circ p and ϖ∘p\varpi\circ p.

The controller regulates the transverse outputs α1,…,αℓ\alpha_1,\ldots,\alpha_\ell and, optionally, the rate of progress along Γ\Gamma. Let r1,…,rℓ≥1r_1,\ldots,r_\ell\ge1 and, when the rate of progress is regulated, r0≥2r_0\ge2 be integers such that, on XUX_U and for j=1,…,mj=1,\ldots,m, LgjLfιαi=0L_{g_j}L_f^\iota\alpha_i=0 for 0≤ι<ri−10\le\iota<r_i-1 and LgjLfιϖ=0L_{g_j}L_f^\iota\varpi=0 for 0≤ι<r0−10\le\iota<r_0-1. The transverse coordinates are ξ=(ξ1,…,ξℓ)∈Rr\xi=(\xi^1,\ldots,\xi^\ell)\in\mathbb{R}^{r} with ξi=(αi,Lfαi,…,Lfri−1αi)\xi^i=(\alpha_i,L_f\alpha_i,\ldots,L_f^{r_i-1}\alpha_i) and r=r1+⋯+rℓr=r_1+\cdots+r_\ell, and the tangential coordinates are η=(ϖ,Lfϖ,…,Lfr0−1ϖ)\eta=(\varpi,L_f\varpi,\ldots,L_f^{r_0-1}\varpi). If the rate of progress is regulated, let k=ℓ+1k=\ell+1 and let the coordinate transformation be T=(η,ξ)\mathcal{T}=(\eta,\xi). Otherwise, let k=ℓk=\ell and T=ξ\mathcal{T}=\xi. In both cases we assume k≤mk\le m. The decoupling matrix D:XU→Rk×mD:X_U\to\mathbb{R}^{k\times m} has the rows (LgjLfr0−1ϖ)j=1m(L_{g_j}L_f^{r_0-1}\varpi)_{j=1}^{m}, in the first case only, and (LgjLfri−1αi)j=1m(L_{g_j}L_f^{r_i-1}\alpha_i)_{j=1}^{m} for i=1,…,ℓi=1,\ldots,\ell. Wherever DD has full row rank, the regulated outputs have vector relative degree (r0,r1,…,rℓ)(r_0,r_1,\ldots,r_\ell), or (r1,…,rℓ)(r_1,\ldots,r_\ell), and the TFL controller is

u=D(x)†((Lfr0ϖLfr1α1⋮Lfrℓαℓ)−(ν∥ν⊥,1⋮ν⊥,ℓ)),ν⊥,i=Λi⊤ξi,ν∥=Λη⊤(η2−ηˉ,η3,…,ηr0),(2)u=D(x)^\dagger\left( \begin{pmatrix} L_f^{r_0}\varpi\\ L_f^{r_1}\alpha_1\\ \vdots\\ L_f^{r_\ell}\alpha_\ell \end{pmatrix} - \begin{pmatrix} \nu_{\parallel}\\ \nu_{\perp,1}\\ \vdots\\ \nu_{\perp,\ell} \end{pmatrix} \right),\qquad \nu_{\perp,i}=\Lambda_i^{\top}\xi^i,\quad \nu_{\parallel}=\Lambda_\eta^{\top}(\eta_2-\bar{\eta},\eta_3,\ldots,\eta_{r_0}), \tag*{(2)}

where the first row of each vector is present only when the rate of progress is regulated, D†=D⊤(DD⊤)−1D^\dagger=D^\top(DD^\top)^{-1} is the right inverse of DD (equal to D−1D^{-1} when k=mk=m), and ηˉ∈R\bar{\eta}\in\mathbb{R} is the reference rate of progress. The error coordinates are z=(ξ,η2−ηˉ,η3,…,ηr0)z=(\xi,\eta_2-\bar{\eta},\eta_3,\ldots,\eta_{r_0}) when the rate of progress is regulated and z=ξz=\xi otherwise, with dimension nz=r+r0−1n_z=r+r_0-1 or nz=rn_z=r. The gains Λi∈Rri\Lambda_i\in\mathbb{R}^{r_i} and Λη∈Rr0−1\Lambda_\eta\in\mathbb{R}^{r_0-1} are chosen such that z˙=Az\dot{z}=Az with AA Hurwitz and block diagonal. The remaining n−nzn-n_z dimensions of the state carry the internal dynamics of the closed loop, which contain the motion along Γ\Gamma ([43]). They may be unstable, and our results do not assume their stability.

Path following manifold. As in Nielsen and Maggiore [43] and Akhtar et al. [7], the path following manifold Γ⋆\Gamma^{\star} is the maximal controlled invariant subset of the lift Γ~={x∈XU:α(x)=0}\widetilde{\Gamma}=\{x\in X_U:\alpha(x)=0\} of the path. In other words, it is the set of states from which a control signal can keep the position on Γ\Gamma for all future time. Wherever the rows of DD that belong to α1,…,αℓ\alpha_1,\ldots,\alpha_\ell are linearly independent, Γ⋆\Gamma^{\star} is the embedded submanifold {ξ=0}\{\xi=0\} of dimension n−rn-r (Appendix § 8).

Examples. Our experiments use two systems that regulate the rate of progress, both with m=k=2m=k=2. The dynamically extended unicycle has n=4n=4, r1=r0=2r_1=r_0=2 and nz=3n_z=3, and it follows a unit circle and a sinusoid. The planar quadrotor with the thrust as a controller state has n=6n=6, r1=r0=3r_1=r_0=3 and nz=5n_z=5, and it follows a vertical unit circle. Appendix § 8 gives their dynamics, gains, thresholds and state constraints.

The admissible set

Let B⊂XB\subset X be a compact box in a fixed chart, with periodic coordinates identified, whose bounds are the state constraints. In this chart, let JTJ_{\mathcal{T}} be the Jacobian of T\mathcal{T}, let ΔT=det⁡(JTJTT)\Delta_{\mathcal{T}}=\sqrt{\det(J_{\mathcal{T}}J_{\mathcal{T}}^{\mathsf{T}})}, and let ΔD=det⁡(DDT)\Delta_D=\sqrt{\det(DD^{\mathsf{T}})}, which equals ∣det⁡D∣|\det D| when k=mk=m. The controller (2) is undefined where DD loses rank, the coordinate transformation T\mathcal{T} loses rank where ΔT=0\Delta_{\mathcal{T}}=0, and the progress map may lose smoothness where its gradient is unbounded, as at the centre of a circular path. We call the union of these sets the singular set. Given thresholds δD,δT,δϖ>0\delta_D,\delta_{\mathcal{T}},\delta_{\varpi}>0, the admissible set is

Ωad={x∈XU∩B∘:ΔD(x)>δD, ΔT(x)>δT, ∥∇ϖ(p(x))∥<δϖ−1},(3)\Omega_{\mathrm{ad}}=\left\{x\in X_U\cap B^{\circ}:\Delta_D(x)>\delta_D,\ \Delta_{\mathcal{T}}(x)>\delta_{\mathcal{T}},\ \lVert\nabla\varpi(p(x))\rVert<\delta_{\varpi}^{-1}\right\}, \tag*{(3)}

where the last inequality is imposed only when the rate of progress is regulated. The closure Ω‾ad\overline{\Omega}_{\mathrm{ad}} is compact because BB is. We assume Ω‾ad⊂XU\overline{\Omega}_{\mathrm{ad}}\subset X_U and call the inequalities in (3) the regularity constraints. Their positive thresholds keep Ω‾ad\overline{\Omega}_{\mathrm{ad}} off the singular set, so the closed-loop vector field fclf_{\mathrm{cl}} of (1) and (2) is smooth on a neighbourhood of Ω‾ad\overline{\Omega}_{\mathrm{ad}}. Since uu scales with 1/ΔD1/\Delta_D, fclf_{\mathrm{cl}} grows large near the singular set, so we integrate the normalized field fˉ=fcl/a\bar{f}=f_{\mathrm{cl}}/a, where a=1+∥Sfcl∥2a=\sqrt{1+\lVert Sf_{\mathrm{cl}}\rVert^2} and SS is the diagonal matrix of inverse half-widths of BB. This change of time scale preserves orbits, forward invariance and convergence within Ωad\Omega_{\mathrm{ad}}, and all step lengths in training and in the tests are measured in the time of fˉ\bar{f}, which we call normalized time.

Target set and constrained region of attraction

The target set is M={x∈Ω‾ad:z(x)=0}M=\{x\in\overline{\Omega}_{\mathrm{ad}}:z(x)=0\}. Since ΔD≥δD>0\Delta_D\ge\delta_D>0 on Ω‾ad\overline{\Omega}_{\mathrm{ad}}, DD has full row rank there, so M⊂Γ⋆M\subset\Gamma^{\star}. Without regulation of the rate of progress, M=Γ⋆∩Ω‾adM=\Gamma^{\star}\cap\overline{\Omega}_{\mathrm{ad}}. With it, MM is the part of Γ⋆∩Ω‾ad\Gamma^{\star}\cap\overline{\Omega}_{\mathrm{ad}} on which the path is traversed at the reference rate. We make the following assumption.

Assumption 1 (a) The closed-loop vector field fclf_{\mathrm{cl}} and the error map zz are smooth on a neighbourhood of Ω‾ad\overline{\Omega}_{\mathrm{ad}}. (b) Lfclz(x)=Az(x)\mathcal{L}_{f_{\mathrm{cl}}}z(x)=Az(x) for all x∈Ωadx\in\Omega_{\mathrm{ad}}, with AA Hurwitz. (c) ∅≠M⊂Ωad\varnothing\ne M\subset\Omega_{\mathrm{ad}}, so z≠0z\ne0 on ∂Ωad\partial\Omega_{\mathrm{ad}}.

Parts (a) and (b) follow from § 2.1 and § 2.2. Part (c) keeps the motion along the path inside the state constraints and off the singular set, and Appendix § 8 checks it for our examples. The set MM is compact and forward invariant (Appendix § 9), and it may be disconnected, as for the unicycle, which can follow the path driving forward or in reverse (Appendix § 8). The constrained region of attraction of MM is

D(M)={x∈Ωad:y(t;x)∈Ωad for all t≥0 and lim⁡t→∞dist⁡(y(t;x),M)=0}.(4)\mathcal{D}(M)=\left\{x\in\Omega_{\mathrm{ad}}:y(t;x)\in\Omega_{\mathrm{ad}}\text{ for all }t\ge0\text{ and }\lim_{t\to\infty}\operatorname{dist}(y(t;x),M)=0\right\}. \tag*{(4)}

By Theorem 2, a trajectory from a state of Ωad∖D(M)\Omega_{\mathrm{ad}} \setminus D(M) leaves Ωad\Omega_{\mathrm{ad}}, by violating a regularity constraint or a state constraint. Deciding membership in D(M)D(M) directly thus requires simulation.

Problem 1 Given fclf_{\mathrm{cl}}, MM and Ωad\Omega_{\mathrm{ad}} satisfying Assumption 1, compute a set Ω\Omega that contains a neighbourhood of MM, is described by sublevel sets of explicit functions, and lies in D(M)D(M) up to a set of bounded measure, with a stated confidence. The set must be obtained without simulating trajectories until they converge, and it should cover as large a share of D(M)D(M) as possible.

Methodology

We address Problem 1 with a two-layer estimate. The inner layer I\mathcal{I} is a sublevel set of a known quadratic function of the error coordinates, and its inclusion in D(M)D(M) is verified by interval branch and bound. The outer layer is a sublevel set of a learned approximation of the state-constrained Zubov value function, and its inclusion is tested on samples. The estimate Ω\Omega is the union of the two layers. This section describes its steps, which Algorithm 1 collects at the end, and § 4 states the guarantees of each layer.

Decomposition of the value function

Let ζ=Szz\zeta= S_z z, where Sz∈Rnz×nzS_z \in\mathbb{R}^{n_z \times n_z} is a positive diagonal scaling matrix. Since A~=SzASz−1\widetilde{A} = S_z A S_z^{-1} is Hurwitz, for every symmetric Q≻0Q \succ0 the Lyapunov equation A~TP+PA~=−Q\widetilde{A}^{\mathsf{T}}P + P\widetilde{A} = -Q has a unique symmetric solution P≻0P \succ0. The quadratic cost-to-go is qex(x)=ζ(x)TPζ(x)q_{\mathrm{ex}}(x) = \zeta(x)^{\mathsf{T}}P\zeta(x), and the running cost is L(x)=γζ(x)TQζ(x)L(x) = \gamma\zeta(x)^{\mathsf{T}}Q\zeta(x) with a cost scale γ>0\gamma> 0. Appendix B gives SzS_z, QQ and γ\gamma for our examples.

Let did_i, i=1,…,Nci = 1,\ldots,N_c, be the distances to the regularity constraints of (3) and to the nonperiodic faces of BB. Each did_i is locally Lipschitz on a neighbourhood of Ωad\Omega_{\mathrm{ad}} and positive on Ωad\Omega_{\mathrm{ad}}, and at every point of ∂Ωad\partial\Omega_{\mathrm{ad}} at least one did_i vanishes. The obstacle function h:XU→[0,∞]h : X_U \to[0,\infty] is

h(x)={∑i=1Ncrelu⁡(bidi(x)−1)2,x∈Ωad,+∞,x∉Ωad,(5)h(x) = \begin{cases} \sum_{i=1}^{N_c} \operatorname{relu}\left(\frac{b_i}{d_i(x)} - 1\right)^2, & x \in\Omega_{\mathrm{ad}},\\ +\infty, & x \notin\Omega_{\mathrm{ad}}, \end{cases} \tag*{(5)}

where the constants bib_i satisfy 0<bi<min⁡x∈Mdi(x)0 < b_i < \min_{x \in M} d_i(x). Hence hh is locally Lipschitz on Ωad\Omega_{\mathrm{ad}}, vanishes on a neighbourhood of MM, and tends to infinity at ∂Ωad\partial\Omega_{\mathrm{ad}}. For a state whose trajectory exists and stays in Ωad\Omega_{\mathrm{ad}} for all t≥0t \ge0, the state-constrained Zubov value function of Grüne and Zidani [24] and its bounded transform are

V(x)=sup⁡t≥0[∫0tL(y(s;x)) ds+h(y(t;x))],v(x)=1−e−V(x),(6)V(x) = \sup_{t \ge0}\left[\int_{0}^{t} L(y(s;x))\,\mathrm{d}s + h(y(t;x))\right], \qquad v(x) = 1 - e^{-V(x)}, \tag*{(6)}

with integration in physical time. For every other state, V(x)=+∞V(x) = +\infty and v(x)=1v(x) = 1. Theorem 2 in § 4 shows that D(M)={x∈Ωad:V(x)<∞}={x∈Ωad:v(x)<1}D(M) = \{x \in\Omega_{\mathrm{ad}} : V(x) < \infty\} = \{x \in\Omega_{\mathrm{ad}} : v(x) < 1\}.

Lemma 1 Let Assumption 1 hold. Then Lfclqex(x)=−ζ(x)TQζ(x)L_{f_{\mathrm{cl}}}q_{\mathrm{ex}}(x) = -\zeta(x)^{\mathsf{T}}Q\zeta(x) for all x∈Ωadx \in\Omega_{\mathrm{ad}}. Consequently, for every x∈Ωadx \in\Omega_{\mathrm{ad}} and t≥0t \ge0 such that y(s;x)∈Ωady(s;x) \in\Omega_{\mathrm{ad}} for all s∈[0,t]s \in[0,t], ∫0tL(y(s;x)) ds=γ(qex(x)−qex(y(t;x))).\int_{0}^{t} L(y(s;x))\,\mathrm{d}s = \gamma(q_{\mathrm{ex}}(x) - q_{\mathrm{ex}}(y(t;x))). Define the margin function ψ:Ωad→[0,∞)\psi: \Omega_{\mathrm{ad}} \to[0,\infty) by ψ=relu⁡(h−γqex)\psi= \operatorname{relu}(h - \gamma q_{\mathrm{ex}}). For every $x \in \Omega_{\mathrm{ad}},

V(x)=γqex(x)+Φ(x),Φ(x)=sup⁡t≥0ψ(y(t;x)),(7)V(x) = \gamma q_{\mathrm{ex}}(x) + \Phi(x), \qquad\Phi(x) = \sup_{t \ge0}\psi(y(t;x)), \tag*{(7)}

where Φ(x)=+∞\Phi(x) = +\infty if the trajectory of xx leaves Ωad\Omega_{\mathrm{ad}}. In particular, V≥γexV \ge\gamma_{\mathrm{ex}} and Φ≥ψ≥0\Phi\ge\psi\ge0 on Ωad\Omega_{\mathrm{ad}}. Moreover, for every τ>0\tau> 0 and every x∈Ωadx \in\Omega_{\mathrm{ad}} such that y(s;x)∈Ωady(s;x) \in\Omega_{\mathrm{ad}} for all s∈[0,τ]s \in[0,\tau],

Φ(x)=max⁡{max⁡0≤s≤τψ(y(s;x)),Φ(y(τ;x))}(8)\Phi(x) = \max\left\{ \max_{0 \le s \le\tau} \psi(y(s;x)), \Phi(y(\tau;x)) \right\} \tag*{(8)}

The proof is in Appendix § 9. Identity (8) has the form of the self-consistency equation of an undiscounted safety value function [55], so we call Φ\Phi the safety value function. Since γex\gamma_{\mathrm{ex}} is known in closed form, only Φ\Phi has to be learned, and the quadratic term supplies the strict decrease of VV along trajectories where Φ\Phi is constant. Moreover, the bound V≥γexV \ge\gamma_{\mathrm{ex}} holds for every approximation of the form γex\gamma_{\mathrm{ex}} plus a nonnegative function, which gives the inclusion of Theorem 3(ii). When the inner layer is not negligible, we add a shaping cost to L\mathcal{L}, to steepen vv just outside the inner layer. Its integral is not known in closed form, so Φ\Phi absorbs it, and the decomposition, the training and the guarantees extend to it (Appendix § 12). For brevity, this section presents the method without it.

Neural candidate

We approximate Φ\Phi on Ωad\Omega_{\mathrm{ad}} in the form

Φθ=max⁡{ψ,Ψθ},Vθ=γex+Φθ,vθ=1−e−Vθ,Ψθ(x)=tanh⁡(qex(x)/q0)1βsoftplus⁡(βNθ(F(x))),(9)\begin{aligned} \Phi_\theta&= \max\{\psi,\Psi_\theta\}, \qquad V_\theta= \gamma_{\mathrm{ex}}+\Phi_\theta, \qquad v_\theta= 1-e^{-V_\theta},\\ \Psi_\theta(x) &= \tanh(q_{\mathrm{ex}}(x)/q_0)\frac{1}{\beta}\operatorname{softplus}(\beta\mathcal{N}_\theta(F(x))), \tag*{(9)} \end{aligned}

with Vθ=+∞V_\theta=+\infty and vθ=1v_\theta=1 outside Ωad\Omega_{\mathrm{ad}}, and we call vθv_\theta the candidate. Here Nθ\mathcal{N}_\theta is a multilayer perceptron with weights θ\theta and β=10\beta=10. Its features F(x)F(x) are built from the state, the error coordinates and the distances to the singular set (Appendix § 12). The constant q0>0q_0>0 is the first percentile of qexq_{\mathrm{ex}} over the collocation points, the admissible samples on which the network is trained (Appendix § 12), and the factor tanh⁡(qex/q0)\tanh(q_{\mathrm{ex}}/q_0) gives Ψθ=0\Psi_\theta=0 on MM. Like the exact solution, VθV_\theta is γex\gamma_{\mathrm{ex}} plus a function that is at least ψ\psi. Hence Vθ≥max⁡{h,γex}V_\theta\ge\max\{h,\gamma_{\mathrm{ex}}\} on Ωad\Omega_{\mathrm{ad}}, and every sublevel set {Vθ<c}\{V_\theta<c\} with c<∞c<\infty lies in {h<c}⊂Ωad\{h<c\}\subset\Omega_{\mathrm{ad}}. These sublevel sets therefore satisfy the regularity constraints and the state constraints by construction, although their forward invariance is not automatic.

Training and level selection

Fitted value iteration. Training solves (8) by regression. The update follows the semi-Lagrangian scheme (5.1) of Grüne and Zidani [24], with a target network as in reach-avoid reinforcement learning [26]. Let yˉ(s;x)\bar{y}(s;x) denote the solution of x˙=fˉ(x)\dot{x}=\bar{f}(x). Since (8) concerns the orbit only, it holds in normalized time and for every step τ>0\tau>0. For each collocation point xx, we integrate fˉ\bar{f} by the fourth-order Runge-Kutta method (RK4) over a step τ\tau with substeps of length 0.0250.025. We then store the end point xτx_\tau and the running maximum Rτ(x)=max⁡0≤s≤τψ(yˉ(s;x))R_\tau(x)=\max_{0\le s\le\tau}\psi(\bar{y}(s;x)) over the substep end points, including s=0s=0, with Rτ(x)=+∞R_\tau(x)=+\infty if a substep leaves Ωad\Omega_{\mathrm{ad}}. These quantities do not depend on the network, so they are computed once. Training keeps an exponential moving average of the weights. With a copy θˉ\bar{\theta} of these averaged weights, refreshed after every epoch, the target is

U(x)=exp⁡(−max⁡{Rτ(x),Φθˉ(xτ)})∈[0,1],(10)U(x)=\exp\left(-\max\{R_\tau(x),\Phi_{\bar{\theta}}(x_\tau)\}\right)\in[0,1], \tag*{(10)}

which is bounded even where the exact value is infinite and equals zero when the step leaves Ωad\Omega_{\mathrm{ad}}. Since Rτ(x)≥ψ(x)R_\tau(x)\ge\psi(x), the loss is (e−Ψθ(x)−U(x))2(e^{-\Psi_\theta(x)}-U(x))^2 where U(x)<e−ψ(x)U(x)<e^{-\psi(x)}, and elsewhere the one-sided loss relu⁡(U(x)−e−Ψθ(x))2\operatorname{relu}(U(x)-e^{-\Psi_\theta(x)})^2 only keeps Ψθ\Psi_\theta below ψ\psi, so that zero loss implies e−Φθ(x)=U(x)e^{-\Phi_\theta(x)}=U(x). Finally, for a point whose step leaves Ωad\Omega_{\mathrm{ad}}, this loss and the obstacle penalty below carry the factor 1+wex1+w_{\mathrm{ex}} with an exit weight wex≥0w_{\mathrm{ex}}\ge0.

Step setting. Unrolling (10) over successive steps shows that its only continuous fixed point that vanishes on MM is sup⁡t≥0ψ(y(t;x))\sup_{t\geq0}\psi(y(t;x)) for every step τ\tau, which is Φ\Phi because the decomposition is exact (Appendix § 12). A long step is therefore admissible on a TFL closed loop, and it carries the value of its end point over a long arc of the orbit in each update. Since the best step within a training budget still depends on the system, the method does not fix it in advance. It trains seed 0 with a long-step setting (τ=2\tau= 2, γ=1/q90\gamma= 1/q_{90}, wex=10w_{\mathrm{ex}} = 10) and with a short-step setting (τ=0.2\tau= 0.2, γ=3/q90\gamma= 3/q_{90}, wex=0w_{\mathrm{ex}} = 0), where q90q_{90} is the 90th percentile of qexq_{\mathrm{ex}} over admissible samples. It then keeps, for all seeds, the setting whose estimate holds the larger share of box samples after the level selection and the final test described below, before fine-tuning. This share is the score by which level selection ranks snapshots, and it uses no reference labels.

Penalties. In addition to the regression loss, three penalties encode properties of the exact solution. The finite-step penalty asks VθV_\theta to decrease over steps of unit length in normalized time that stay in Ωad\Omega_{\mathrm{ad}}, the decrease penalty asks for LfclΦθ≤0L_{f_{\mathrm{cl}}}\Phi_\theta\leq0 where Ψθ>ψ\Psi_\theta> \psi, and the obstacle penalty asks for Ψθ>ψ\Psi_\theta> \psi where Lfclψ>0L_{f_{\mathrm{cl}}}\psi> 0. Except on the quadrotor, every term carries an exponential loss weight, which keeps the gradient at the levels near one where the outer level is decided. In addition, the states whose unit step violates the decrease most join the data during training (Appendix § 12).

Level selection. The inner layer is I={x∈B∘∩XU:qex(x)<cI}\mathcal{I} = \{x \in B^\circ\cap X_U : q_{\mathrm{ex}}(x) < c_I\}, with cIc_I one percent below the largest tested level that passes the interval verification of § 4, and c1=1−e−cIc_1 = 1 - e^{-c_I}. A state x∈Ωadx \in\Omega_{\mathrm{ad}} satisfies the finite-step decrease condition for a function v~:Ωad→[0,1)\tilde{v} : \Omega_{\mathrm{ad}} \to[0,1) if, for some horizon H∈{1,2,4,8}H \in\{1,2,4,8\}, yˉ(s;x)∈Ωad\bar{y}(s;x) \in\Omega_{\mathrm{ad}} for all s∈[0,H]s \in[0,H] and v~(yˉ(H;x))<v~(x)\tilde{v}(\bar{y}(H;x)) < \tilde{v}(x). In level selection, v~=vθ\tilde{v} = v_\theta. A state with vθ(x)≥c1v_\theta(x) \geq c_1 that fails this condition is a counterexample. The condition is tested numerically, with the exemptions of Appendix § 12. The outer level c2c_2 is the lowest value of vθv_\theta among the sampled counterexamples, or one if none is found, and we call {c1≤vθ<c2}\{c_1 \leq v_\theta< c_2\} the annulus. Training saves a snapshot of the network every five epochs. Screening sets c2c_2 for every snapshot from counterexamples sought on samples of the box, of its faces and next to the singular set, and ranks the snapshots by the share of box samples in I∪{vθ<c2}\mathcal{I} \cup\{v_\theta< c_2\}. The three best are then retested on fresh samples, and the best one after the retest is kept (Appendix § 12).

Counterexample-guided fine-tuning. Since the outer level is the lowest value among the counterexamples, a handful of states determines it. Each of four rounds of fine-tuning therefore searches fresh samples for counterexamples with vθ(x)≤c2+0.05v_\theta(x) \leq c_2 + 0.05 and runs 300 Adam steps on them, with a decrease margin and an anchor term that holds Ψθ\Psi_\theta elsewhere. The networks of these rounds and the trained network are screened like training snapshots. With the kept network and its level cˉ2\bar{c}_2 fixed, the final test draws N=106N = 10^6 i.i.d. uniform samples of the box, with a seed disjoint from those of all preceding stages. The final level c2c_2 is then the minimum of cˉ2\bar{c}_2 and the values of vθv_\theta at the counterexamples that this test finds. If the kept network is the trained one, the counterexamples of its earlier final test count as well. Algorithm 1 summarizes the method.

Guarantees

This section characterizes D(M)\mathcal{D}(M), gives sufficient conditions for the inclusion of both layers of the estimate Ω=I∪{vθ<c2}\Omega= \mathcal{I} \cup\{v_\theta< c_2\} in D(M)\mathcal{D}(M), and states what the tests of the outer layer guarantee. The proofs are in Appendix § 9.

Theorem 2 Let Assumption 1 hold. A state x∈Ωadx \in\Omega_{\mathrm{ad}} belongs to D(M)\mathcal{D}(M) if and only if its trajectory exists and remains in Ωad\Omega_{\mathrm{ad}} for all t≥0t \geq0. Consequently, D(M)\mathcal{D}(M) is the largest forward invariant subset of Ωad\Omega_{\mathrm{ad}}, and D(M)={x∈Ωad:V(x)<∞}={x∈Ωad:v(x)<1}\mathcal{D}(M) = \{x \in\Omega_{\mathrm{ad}} : V(x) < \infty\} = \{x \in\Omega_{\mathrm{ad}} : v(x) < 1\}.

The set equality is the conclusion of Grüne and Zidani [24], here for a target set that may be disconnected and without stability assumptions on the internal dynamics. The next result gives sufficient conditions for the inclusion of both layers in D(M)\mathcal{D}(M).

Algorithm 1: Two-layer estimate of the constrained region of attraction of a TFL closed loop
Input: closed loop fclf_{\mathrm{cl}}, admissible set Ωad\Omega_{\mathrm{ad}}, quadratic cost-to-go qexq_{\mathrm{ex}}, obstacle function hh, set S\mathcal{S} of step settings (τ,γ,wex)(\tau,\gamma,w_{\mathrm{ex}})
Output: estimate Ω=I∪{vθ<c2}\Omega=\mathcal{I}\cup\{v_\theta<c_2\}
Inner layer: set cIc_{\mathcal{I}} one percent below the largest tested level that passes the interval verification of § 4, and set I={γqex<cI}\mathcal{I}=\{\gamma q_{\mathrm{ex}}<c_{\mathcal{I}}\}, c1=1−e−cIc_1=1-e^{-c_{\mathcal{I}}}
Step setting: on seed 0, run the lines below up to the level selection and a final test for every setting in S\mathcal{S}, and keep, for all seeds, the setting whose estimate holds the largest share of box samples
Step data: from every collocation point xx, integrate fˉ\bar{f} over the step τ\tau and store the end point xτx_\tau and the running maximum Rτ(x)R_\tau(x)
for each epoch do
freeze a copy θˉ\bar{\theta} of the averaged weights and form the target UU of (10)
update θ\theta on the regression loss of Φθ=max⁡{ψ,Ψθ}\Phi_\theta=\max\{\psi,\Psi_\theta\} to UU, with the factor 1+wex1+w_{\mathrm{ex}} on points whose step leaves Ωad\Omega_{\mathrm{ad}}, plus the three penalties
add the states that violate the finite-step decrease most to the data
end
Level selection: for every snapshot, set c2c_2 to the lowest value of vθ=1−e−(γqex+Φθ)v_\theta=1-e^{-(\gamma q_{\mathrm{ex}}+\Phi_\theta)} among the sampled counterexamples, and keep the snapshot whose estimate holds the largest share of box samples
Fine-tuning: for four rounds, search fresh samples for counterexamples with vθ≤c2+0.05v_\theta\le c_2+0.05, train on them and repeat the level selection
Final test: draw N=106N=10^6 i.i.d. samples and lower c2c_2 to the lowest value of vθv_\theta among the counterexamples found

Table 1.

Theorem 3 Let Assumption 1 hold, let cI>0c_{\mathcal{I}}>0, I={x∈B∘∩XU:γqex(x)<cI}\mathcal{I}=\{x\in B^\circ\cap X_U:\gamma q_{\mathrm{ex}}(x)<c_{\mathcal{I}}\} and c1=1−e−cIc_1=1-e^{-c_{\mathcal{I}}}. (i) Suppose that every x∈B∩XUx\in B\cap X_U with γqex(x)<cI\gamma q_{\mathrm{ex}}(x)<c_{\mathcal{I}} satisfies the regularity constraints of (3), and that at every such point on ∂B\partial B the field fclf_{\mathrm{cl}} points strictly into BB across every active face. Then I\mathcal{I} is forward invariant and M⊂I⊂D(M)M\subset\mathcal{I}\subset\mathcal{D}(M). (ii) Let v~:Ωad→[0,1)\tilde{v}:\Omega_{\mathrm{ad}}\to[0,1) satisfy −ln⁡(1−v~)≥γqex-\ln(1-\tilde{v})\ge\gamma q_{\mathrm{ex}}, as vθv_\theta does for every θ\theta. Then {v~<c1}⊂I\{\tilde{v}<c_1\}\subset\mathcal{I}. (iii) Let (i) hold, let v~\tilde{v} be as in (ii) and let c2∈[c1,1]c_2\in[c_1,1]. Suppose that every x∈Ωad∖Ix\in\Omega_{\mathrm{ad}}\setminus\mathcal{I} with c1≤v~(x)<c2c_1\le\tilde{v}(x)<c_2 satisfies the finite-step decrease condition for v~\tilde{v}. Then {v~<c2}⊂D(M)\{\tilde{v}<c_2\}\subset\mathcal{D}(M).

The inclusion {vθ<c1}⊂I\{v_\theta<c_1\}\subset\mathcal{I} of part (ii) holds by construction for every candidate, and we check the hypotheses of part (i) by interval branch and bound (Moore et al., [41]) in floating-point arithmetic with outward rounding (Appendix § 11).

Statistical bounds. Two bounds quantify what the sample tests can miss (Appendix § 9). First, the final test fixes the network and the level cˉ2\bar{c}_2 before it draws its NN samples, so a scenario bound applies to it (Proposition 4). With probability at least 1−(1−ε)N1-(1-\varepsilon)^N, the counterexamples of the implemented test that remain in the annulus have measure at most ε\varepsilon. However, such a bound does not bound the measure of the states of the outer layer outside D(M)\mathcal{D}(M) (Kenanian et al., [29]). Second, the direct test of unsafe states therefore simulates NN fresh i.i.d. samples until they reach the inner layer and counts the KK samples of an estimate that fail to do so inside Ωad\Omega_{\mathrm{ad}}. With confidence 1−10−61-10^{-6}, the states of the estimate outside D(M)\mathcal{D}(M) then have measure at most the one-sided Clopper–Pearson bound εK\varepsilon_K (Proposition 5), which is about ln⁡(106)/N\ln(10^6)/N for K=0K=0.

CaseShare of boxInnerQuadraticZubov PINNDirichlet-Zubov PINNNo decomp.OursExact vvFP
Circle84.86.48.36.4±0.06.4 \pm0.011.3±3.911.3 \pm3.911.8±3.211.8 \pm3.272.0±1.672.0 \pm1.685.40.00
Sinusoid95.01.12.14.0±0.64.0 \pm0.61.1±0.01.1 \pm0.031.4±20.931.4 \pm20.983.4±0.883.4 \pm0.888.10.00
Quadrotor56.20.00.00.0±0.00.0 \pm0.00.0±0.00.0 \pm0.00.0±0.00.0 \pm0.019.1±4.019.1 \pm4.046.80.00

Table 1 (PDF p. 9). Coverage (%) of the simulated region of attraction on the TFL cases, mean ±\pm std over four seeds. Exact vv: exact value function cut at the level of the candidate. FP: largest share of reference states inside an estimate and outside that region.

What is not proved. The interval verification of Theorem 3(i) relies on the accuracy of the mathematical library and on the one percent reduction of the level, so we call the inner layer verified, not proved. The hypothesis of part (iii) is tested on samples only, so the outer layer is not proved either. The direct test classifies each sample by a finite simulation, so its bound holds up to the integration error (Appendix G). It changes no estimate, so the estimate itself is still obtained without simulating trajectories until they converge.

Simulations

Setup. Tables 5 and 6 in Appendix § 12 list the configuration. It is the same for all systems, except for two loss weights and the shaping cost on the quadrotor (Appendix § 12). Four seeds are trained per case. The reference simulation classifies 5×1045 \times10^{4} quasi-random states of the box by integrating the closed loop until the state converges to MM or leaves Ωad\Omega_{\mathrm{ad}} (Appendix § 12). The simulated regions of attraction hold 84.8 %, 95.0 % and 56.2 % of the three boxes (Figure 1). The coverage of an estimate is the share of the reference states in the simulated region of attraction that it contains. Training, the step setting and level selection never use the reference simulation. However, we developed the configuration, including the two step settings, while watching coverage on seed 0 (Appendix § 12), and all reported runs, seed 0 included, were trained after these choices. The step setting keeps the long step on all three cases.

Figure 1. Slices of the three cases, the unicycle at heading north and speed 1 and the quadrotor at the orbit velocity, level tilt and hover thrust. The estimate is the union over four seeds.

Baselines. All baselines pass through level selection and the final test. The quadratic baseline is 1−e−γqex1 - e^{-\gamma q_{\mathrm{ex}}} alone. The Zubov PINN of Liu et al. [34] and the Dirichlet-Zubov PINN of Meng and Liu [37] learn from simulated labels over four seeds. The scale of their tanh transform is searched on seed 0, and neither receives fine-tuning (Appendix D). The no-decomposition baseline uses the same network, data, losses, level selection and fine-tuning without the term γqex\gamma q_{\mathrm{ex}} in the value function. The safety-value baseline sets γ=0\gamma= 0, so it has no running cost and learns sup⁡th\sup_{t} h along trajectories alone, on seed 0 without fine-tuning.

How large are the estimates? The estimates cover most of the simulated region of attraction on the unicycle and a smaller part of it on the quadrotor (Table 1). Fine-tuning raises the mean single-seed coverage from 70.1 % to 72.0 % on the circle and from 82.1 % to 83.4 % on the sinusoid. No reference state lies inside an estimate and outside the simulated region of attraction. The direct test of unsafe states, with 10610^{6} i.i.d. box samples per case, finds one unsafe sample in the union of four estimates of each case. With confidence 1−10−61 - 10^{-6} and up to the integration error, the states of each union outside D(M)\mathcal{D}(M) therefore occupy at most 1.7×10−51.7 \times10^{-5} of the box (Appendix G). The quadrotor is the hardest of the three cases, with six states and a negligible inner layer (Appendix G). Counterexamples hold its outer levels between 0.12 and 0.25. At these levels the exact value function covers 46.8 %, of which the outer layers reach 41 %, and the bound Vθ≥γqexV_{\theta} \ge\gamma q_{\mathrm{ex}} alone excludes 37.5 % of its simulated region of attraction. Both the fit and the level therefore limit this estimate.

Does the decomposition matter? The candidate covers 60.2, 52.0 and 19.1 points more than the no-decomposition baseline, which must learn the integral of the running cost and covers 11.8 %, 31.4 % and 0.0 % under the same tests. The safety-value baseline has no running cost. On seed 0, before fine-tuning, it covers 11.1 %, 33.1 % and 1.3 %, against 72.0 %, 82.5 % and 24.9 % for the candidate (Table 8 in Appendix G).

CaseVariantc2c_2CoverageFP
Circlefull method0.86972.00.000
CirclePDE residual loss alone0.19510.50.000
Circlesafety-value baseline0.74311.10.002
Circleno finite-step penalty, no counterexamples0.76863.40.000
Circleno exponential loss weight0.77766.70.000
Circlesum form Φ=ψ+Ψθ\Phi= \psi+ \Psi_\theta0.9959.90.000
Sinusoidfull method0.90782.50.000
Sinusoidsafety-value baseline0.92833.10.002
Quadrotorfull method0.31424.90.000
Quadrotorsafety-value baseline0.4831.30.000

Table 8. Ablations on seed 0 before fine-tuning, with the same level selection and final test. FP: share (%) of reference states inside the estimate and outside the simulated region of attraction.

How do the PINN baselines compare? They cover much less than the candidate on all three cases (Table 1), although both use simulated labels and the candidate uses none. The lowest sampled value outside the inner layer belongs to a counterexample in every seed for the Zubov PINN on the circle and the quadrotor, and for the Dirichlet-Zubov PINN on the sinusoid and the quadrotor. In these cases the estimate reduces to the inner layer. However, level selection is not what limits them. Even at the level that the reference simulation itself would allow, no PINN network of any scale, seed, network size or training snapshot covers more than 20.2 %, 33.7 % and 2.4 % (Appendix D).

Closed loops with an equilibrium. On seed 0 of five benchmarks of the literature, the step setting of Algorithm 1 keeps the long step on two and the short step on the others. After fine-tuning, the estimates cover between 86.5 % and 98.9 % of the simulated regions of attraction (Table 3 in Appendix D).

SystemShare, longShare, shortKeptCoverage, longCoverage, shortCoverage, kept, fine-tuned
Circle61.159.3long72.069.874.0
Sinusoid78.454.3long82.557.283.7
Quadrotor14.05.3long24.99.422.9
Van der Pol, μ=1\mu= 130.338.8short77.398.998.9
Van der Pol, μ=3\mu= 315.421.3short54.174.989.2
Pendulum30.227.4long93.584.694.2
3-D system50.040.0long84.767.886.5
Power system15.729.9short47.590.392.4

Table 3. Step setting of Algorithm 1 on seed 0. Share: share of box samples in the estimate (%) before fine-tuning, which decides the setting without reference labels. Coverage (%) of the simulated region of attraction is shown for comparison, before fine-tuning for both settings and after fine-tuning for the kept one.

Conclusions

For a TFL controller with state constraints, the exact linear error dynamics make the constrained region of attraction the largest forward invariant subset of the admissible set. They also split its Zubov value function into a known quadratic cost-to-go and a safety value function. Learning only the latter by fitted value iteration gives a two-layer estimate, with an inner layer verified by interval branch and bound and an outer layer tested on samples. On average, this estimate covers 72.0 %, 83.4 % and 19.1 % of the simulated regions of attraction of a unicycle on two paths and of a planar quadrotor. The method needs the model, and we did not attempt the spatial quadrotor. Two further directions remain. First, a rigorous flow enclosure on the annulus would turn the test of the outer layer into a verification. Second, on the quadrotor the estimate reaches 41 % of what the exact value function covers at the same levels, so both a better fit and higher outer levels remain to be gained.

Generative AI Use: Generative AI tools were used to improve the syntax and grammar of some paragraphs in the manuscript. The authors reviewed and verified all AI-assisted content.

Related Work

Zubov functions with constraints. Camilli et al. [18] extended Zubov’s equation to perturbed systems, and Grüne and Zidani [24] added state constraints through an obstacle term. Our value function is their equation (3.4) without the perturbation. Xue et al. [53] use the same kind of value function with a prescribed running cost built from a polynomial, whose integral along trajectories is unknown, and compute convex sum-of-squares inner approximations for polynomial systems. Lemma 1 gives this integral in closed form for TFL closed loops. Converse Lyapunov-barrier theorems show that a smooth Lyapunov function exists on sets such as D(M)\mathcal{D}(M) when the target set is compact and uniformly asymptotically stable, for perturbed continuous-time systems (Meng et al. [38]) and for hybrid systems (Meng and Liu [36]). Such a function is not known in closed form. Neural solvers of Zubov equations (Liu et al. [34, 33]; Kang et al. [28]) and their state-constrained versions in continuous time (Meng and Liu [37]) and discrete time (Serry et al. [48, 49]) learn the value function of an equilibrium or of a compact invariant set. Those that certify the result verify Lie-derivative conditions in continuous time or one-step conditions in discrete time, with satisfiability-modulo-theories or bound-propagation tools. Xue et al. [52] characterize the interior of the maximal robust region of attraction of an equilibrium of a state-constrained, perturbed discrete-time system by a Bellman equation, a discrete-time analogue of the equation of Grüne and Zidani [24]. Serry et al. [49] note that initial estimates of the region of attraction for non-singleton sets remain open, and our inner layer gives one for TFL closed loops.

Safety value functions and reachability. Equation (8) is the self-consistency equation of an undiscounted safety value function (Yu et al. [55]), and reach-avoid reinforcement learning uses a discounted form of this recursion with a target network (Hsu et al. [26]). Our update follows the semi-Lagrangian scheme (5.1) of Grüne and Zidani [24] with such a network. In the same way, Li et al. [31] bootstrap the data term of a Zubov PINN through the network at the end of a short integration. DeepReach solves reachability equations by a residual loss (Bansal and Tomlin, [12]), and Bokanowski et al. [17] use neural backward recursions for finite-horizon problems with a maximum running cost and obstacles. The closest precedent for our problem is Oishi et al. [45], who synthesize, on a grid and in low dimension, parameterized feedback-linearizing controllers together with the set of states on which they satisfy state and input constraints. Our setting differs in the target set, the singular decoupling matrix, the learned function, and the known decrease that the running cost provides. In the language of viability theory, D(M)\mathcal{D}(M) is the viability kernel of Ωad\Omega_{\mathrm{ad}} under the closed loop.

Local estimates and sampled certificates. Manchester [35] computes inner estimates of the regions of attraction of hybrid limit cycles, starting from a quadratic form on the linearized transverse dynamics and enlarging it by sum-of-squares. Faulwasser and Findeisen [22] use a sublevel set of a quadratic in transverse coordinates as the terminal region of a path-following predictive controller, for a transverse feedback linearization without a singular decoupling matrix. Finite-step Lyapunov conditions go back to Aeyels and Peuteman [2] and Geiselhart et al. [23]. Bobiti and Lazar [15] verify a decrease condition on samples with Lipschitz margins on an annulus around a local set obtained from the linearization. Our annulus has the same shape and is tested on samples only, and our inner layer is verified by interval arithmetic. Liu et al. [34] combine a local quadratic region with a neural sublevel set, with the inclusion checked by a formal verifier. Li et al. [31] certify an annulus above an inner level, and Zanotta et al. [56] certify an annulus statistically above an independently certified inner set. The anchored networks of Liu [32] give an inclusion of sublevel sets with a constant that depends on the parameters. Inclusion (ii) of Theorem 3 holds with no such constant. Boffi et al. [16] bound the generalization error of certificates learned from trajectories, and Proposition 4 is the one-dimensional case of Campi and Garatti [19]. Kenanian et al. [29] show by example that a bound on the measure of counterexamples to the decrease does not bound the measure of non-converging states. Berkenkamp et al. [14] and Richards et al. [46] close the gap between samples and the continuum with Lipschitz constants on a grid. In our setting the Lipschitz constant grows as the distances to the singular set shrink, which makes such a grid expensive in four to six dimensions.

Neural Lyapunov functions and TFL. Neural Lyapunov functions with counterexample-guided training and exact verification ([20, 1, 21, 51, 54]) reach systems of six to twelve states. Barreau and Bastianello [13] learn a maximal Lyapunov function as a quadratic plus a signed residual, optimized jointly on a prescribed compact domain. Our candidate shares this additive form. Its quadratic is fixed and exact on the admissible set, and its network is nonnegative. We found no prior network composed with an obstacle function by a maximum. The TFL literature ([43, 44, 4, 7, 5, 10]) proves local exponential stability of the target set and treats regularity as an assumption. For the quadrotor, Assumption A1 of Al Lawati and Akhtar [10] assumes nonzero thrust on a neighbourhood of the path, and their Remark 2 notes that positive thrust is guaranteed only near the path following manifold.

The Two Systems

Both systems are instances of § 2.1 with ℓ=1\ell= 1, a regulated rate of progress and k=m=2k = m = 2, so D†=D−1D^{\dagger} = D^{-1}, and in both the internal dynamics are the progress ϖ\varpi along the path. Derivatives of circle-valued maps are computed with local angular lifts. The formulation also covers controllers that do not regulate the rate of progress. For example, the kinematic unicycle with constant speed of Nielsen and Maggiore [43] has n=3n = 3, m=1m = 1 and one transverse output with r1=2r_1 = 2, so k=1k = 1, nz=2n_z = 2, and the motion along the path at the fixed speed is internal dynamics.

Path following manifold. Let D⊥:XU→Rℓ×mD_{\perp}: X_U \to\mathbb{R}^{\ell\times m} be the submatrix of DD formed by its rows for α1,…,αℓ\alpha_1,\ldots,\alpha_\ell, and let X⊥={x∈XU:rank⁡D⊥(x)=ℓ}X_{\perp} = \{x \in X_U : \operatorname{rank} D_{\perp}(x) = \ell\}, an open set. On X⊥X_{\perp} the differentials of the components of ξ\xi are linearly independent, by the argument of Isidori [27] with D⊥D_{\perp} of full row rank, so {x∈X⊥:ξ(x)=0}\{x \in X_{\perp} : \xi(x) = 0\} is an embedded submanifold of X⊥X_{\perp} of dimension n−rn-r. It is the maximal connected invariant subset of Γ∩X⊥\Gamma\cap X_{\perp}. Indeed, a state of Γ∩X⊥\Gamma\cap X_{\perp} with ξ≠0\xi\ne0 has a nonzero derivative LflαiL_f^l \alpha_i, l<ril < r_i, that no input affects, so its position leaves Γ\Gamma immediately. Conversely, from a state with ξ=0\xi= 0, the feedback u=−D⊥†(Lfr1α1,…,Lfrℓαℓ)u = -D_{\perp}^{\dagger}(L_f^{r_1}\alpha_1,\ldots,L_f^{r_\ell}\alpha_\ell) keeps ξ=0\xi= 0 while the state remains in X⊥X_{\perp}. Hence Γ∩X⊥={x∈X⊥:ξ(x)=0}\Gamma\cap X_{\perp} = \{x \in X_{\perp} : \xi(x) = 0\}, and a trajectory that stays in a compact subset of X⊥X_{\perp} with ξ→0\xi\to0 converges to Γ\Gamma. Since ΔD>0\Delta_D > 0 implies that DD, and hence D⊥D_{\perp}, has full row rank, the target set of § 2.3 satisfies M⊂ΓM \subset\Gamma.

Unicycle. The dynamically extended unicycle, with n=4n = 4, r1=r0=2r_1 = r_0 = 2 and nz=3n_z = 3, a simpler model than the car-like robot of Akhtar et al. [7] with the same singularity at zero speed, has state x=(x1,x2,x3,x4)x=(x_1,x_2,x_3,x_4) and

x˙1=(1+x4)cos⁡x3,x˙2=(1+x4)sin⁡x3,x˙3=u2,x˙4=u1,(11)\dot{x}_1=(1+x_4)\cos x_3,\quad\dot{x}_2=(1+x_4)\sin x_3,\quad\dot{x}_3=u_2,\quad\dot{x}_4=u_1, \tag*{(11)}

so 1+x41+x_4 is the speed, u1u_1 the acceleration and u2u_2 the turn rate. Both outputs have relative degree two, ξ=(α,Lfα)\xi=(\alpha,L_f\alpha), η=(ϖ,Lfϖ)\eta=(\varpi,L_f\varpi), T(x)=(η,ξ)\mathcal{T}(x)=(\eta,\xi), and

D(x)=[Lg1LfϖLg2LfϖLg1LfαLg2Lfα],det⁡D(x)=±(1+x4)det⁡[∇ϖ∇α].(12)D(x)=\begin{bmatrix}L_{g_1}L_f\varpi& L_{g_2}L_f\varpi\\L_{g_1}L_f\alpha& L_{g_2}L_f\alpha\end{bmatrix},\quad\det D(x)=\pm(1+x_4)\det\begin{bmatrix}\nabla\varpi\\\nabla\alpha\end{bmatrix}. \tag*{(12)}

The controller is (2) with ν⊥=−ξ1−2ξ2\nu_\perp=-\xi_1-2\xi_2 (double pole at −1-1) and ν∥=−(η2−1)\nu_\parallel=-(\eta_2-1), so the rate of progress η2=Lfϖ\eta_2=L_f\varpi is driven to ηˉ=1\bar{\eta}=1. The error coordinates are z=(ξ1,ξ2,η2−1)z=(\xi_1,\xi_2,\eta_2-1) with A=blkdiag⁡([01−1−2],−1)A=\operatorname{blkdiag}\left(\begin{bmatrix}0&1\\-1&-2\end{bmatrix},-1\right) and Q=diag⁡(1.742,0.249,3.281)Q=\operatorname{diag}(1.742,0.249,3.281). The target set M={z=0}M=\{z=0\} is the path traversed at the rate of progress η2=1\eta_2=1. For the unicycle it has two components, because η2\eta_2 depends on the velocity and not on the heading. In the first the heading is along the path and the speed is positive, and in the second the heading is against the path and the speed is negative. On both components, ∣det⁡D∣|\det D| and ∣det⁡JT∣|\det J_{\mathcal{T}}| are bounded away from zero and the state is inside the box, as the expressions below show, so MM is compact and lies in Ωad\Omega_{\mathrm{ad}}. The reference simulation counts convergence to either component.

Circle. The path is α=x12+x22−1\alpha=x_1^2+x_2^2-1 with ϖ=atan2⁡(x2,x1)\varpi=\operatorname{atan2}(x_2,x_1), so ξ2=2(1+x4)(x1cos⁡x3+x2sin⁡x3)\xi_2=2(1+x_4)(x_1\cos x_3+x_2\sin x_3), η2=(1+x4)(x1sin⁡x3−x2cos⁡x3)/(x12+x22)\eta_2=(1+x_4)(x_1\sin x_3-x_2\cos x_3)/(x_1^2+x_2^2), det⁡D=−2(1+x4)\det D=-2(1+x_4), det⁡JT=4(1+x4)\det J_{\mathcal{T}}=4(1+x_4), ∥∇ϖ∥−1=x12+x22\|\nabla\varpi\|^{-1}=\sqrt{x_1^2+x_2^2}. Box: ∣x1∣,∣x2∣≤2|x_1|,|x_2|\leq2, x4∈[−3,1]x_4\in[-3,1] (speed in [−2,2][-2,2]), x3∈S1x_3\in S^1.

Sinusoid. The path is α=x2−sin⁡x1\alpha=x_2-\sin x_1 with ϖ=x1\varpi=x_1, so ξ2=(1+x4)(sin⁡x3−cos⁡x1cos⁡x3)\xi_2=(1+x_4)(\sin x_3-\cos x_1\cos x_3), η2=(1+x4)cos⁡x3\eta_2=(1+x_4)\cos x_3, det⁡D=det⁡JT=1+x4\det D=\det J_{\mathcal{T}}=1+x_4, ∥∇ϖ∥−1=1\|\nabla\varpi\|^{-1}=1. The closed loop is 2π2\pi-periodic in x1x_1, taken in [−π,π)[-\pi,\pi). The box is ∣x2∣≤2|x_2|\leq2, x4∈[−3,1]x_4\in[-3,1]. On MM the speed has magnitude 1+cos⁡2x1≤2\sqrt{1+\cos^2x_1}\leq\sqrt{2}. The thresholds are δD=δT=10−2\delta_D=\delta_{\mathcal{T}}=10^{-2} and δϖ=5⋅10−2\delta_\varpi=5\cdot10^{-2}, so Ωad\Omega_{\mathrm{ad}} excludes ∣1+x4∣≤0.005|1+x_4|\leq0.005 (circle) or ≤0.01\leq0.01 (sinusoid) and, for the circle, the disc of radius 0.050.05 around the centre.

Planar quadrotor. The planar quadrotor, with n=6n=6, r1=r0=3r_1=r_0=3 and nz=5n_z=5, has state (p1,p2,p˙1,p˙2,Θ,T)(p_1,p_2,\dot{p}_1,\dot{p}_2,\Theta,T) and

p¨=T[−sin⁡Θcos⁡Θ]−[01],Θ˙=u2,T˙=u1,(13)\ddot{p}=T\begin{bmatrix}-\sin\Theta\\\cos\Theta\end{bmatrix}-\begin{bmatrix}0\\1\end{bmatrix},\qquad\dot{\Theta}=u_2,\qquad\dot{T}=u_1, \tag*{(13)}

in units with unit mass, unit gravity and unit path radius. The thrust TT is a controller state (the dynamic extension of Akhtar et al. [5]), and the tilt rate is the second input, as provided by a multirotor autopilot with an inner rate loop. For the loop α=p12+p22−1\alpha=p_1^2+p_2^2-1, ϖ=atan2⁡(p2,p1)\varpi=\operatorname{atan2}(p_2,p_1) both outputs have relative degree three: ξ=(α,2pTp˙,2∣p˙∣2+2pTp¨)\xi=(\alpha,2p^{\mathsf{T}}\dot{p},2|\dot{p}|^2+2p^{\mathsf{T}}\ddot{p}), η2=(p1p˙2−p2p˙1)/∣p∣2\eta_2=(p_1\dot{p}_2-p_2\dot{p}_1)/|p|^2 the angular rate, η3=Lf2ϖ\eta_3=L_f^2\varpi, and

D(x)=[ν1/∣p∣2Tν2/∣p∣22ν2−2Tν1],ν1=p1cos⁡Θ+p2sin⁡Θ,ν2=p2cos⁡Θ−p1sin⁡Θ,(14)D(x)=\begin{bmatrix}\nu_1/|p|^2&T\nu_2/|p|^2\\2\nu_2&-2T\nu_1\end{bmatrix},\qquad\nu_1=p_1\cos\Theta+p_2\sin\Theta,\quad\nu_2=p_2\cos\Theta-p_1\sin\Theta, \tag*{(14)}

so det⁡D=−2T\det D=-2T and det⁡JT=−8T\det J_{\mathcal{T}}=-8T. The controller is (2) with ν⊥=−8ξ1−12ξ2−6ξ3\nu_\perp=-8\xi_1-12\xi_2-6\xi_3 (triple pole at −2-2) and ν∥=−(η2−12)−2η3\nu_\parallel=-(\eta_2-\frac{1}{2})-2\eta_3 (double pole at −1-1). The error coordinates are z=(ξ1,ξ2,ξ3,η2−12,η3)z=(\xi_1,\xi_2,\xi_3,\eta_2-\frac{1}{2},\eta_3) with Q=IQ=I, and M={z=0}M=\{z=0\} is the loop traversed counter-clockwise at angular rate 12\frac{1}{2}. On MM the thrust is T=∥p¨+(0,1)∥∈[0.75,1.25]T = \lVert\ddot{p} + (0,1)\rVert\in[0.75, 1.25]. The box is ∣pi∣≤2\lvert p_i\rvert\le2, ∣p˙i∣≤2\lvert\dot{p}_i\rvert\le2, T∈[0,4]T \in[0,4] (thrust-to-weight ratio up to four), Θ∈S1\Theta\in S^1. The thresholds are those of the unicycle, so Ωad\Omega_{\mathrm{ad}} excludes T≤0.005T \le0.005 and the disc of radius 0.05 around the centre. The reference simulation treats a sign change of TT between two steps as a departure through the singular set.

Constants. Table 2 lists the scaling SzS_z and the cost scale γ\gamma. The scale is γ=1/q90\gamma= 1/q_{90} for the long-step setting and γ=3/q90\gamma= 3/q_{90} for the short-step setting of § 3.3, where q90q_{90} is the 90th percentile of qexq_{\mathrm{ex}} over admissible samples. Both constants are computed from the model, without simulating trajectories.

Caseγ\gammaSz−1S_z^{-1}
Circle0.04441.550, 1.396, 1.011
Sinusoid0.06081.001, 0.625, 1.000
Quadrotor0.02881.546, 2.439, 5.953, 0.703, 1.034

Table 2. Constants computed before training: the cost scale γ\gamma of the long-step setting and the diagonal of Sz−1S_z^{-1}.

Proofs

The target set. Since Ω‾ad\overline{\Omega}_{\mathrm{ad}} is compact and zz is continuous, MM is compact. A trajectory from MM keeps z=0z = 0 while it stays in Ωad\Omega_{\mathrm{ad}}. It cannot reach ∂Ωad\partial\Omega_{\mathrm{ad}}, because the first such point would lie in M⊂ΩadM \subset\Omega_{\mathrm{ad}} by continuity of zz and Assumption 1(c). Hence MM is forward invariant.

Proof of Lemma 1. We prove the lemma with the shaping cost of Appendix § 12, for which Ct(x)=κ∫0tρ(y(s;x)) dsC_t(x) = \kappa\int_0^t \rho(y(s;x))\,\mathrm{d}s, and κ=0\kappa= 0 gives the statement. By Assumption 1(b), z˙=Az\dot{z} = Az on Ωad\Omega_{\mathrm{ad}}, hence ζ˙=Szz˙=A~ζ\dot{\zeta} = S_z\dot{z} = \tilde{A}\zeta and Lfclqex=ζT(A~TP+PA~)ζ=−ζTQζL_{f_{\mathrm{cl}}}q_{\mathrm{ex}} = \zeta^{\mathsf{T}}(\tilde{A}^{\mathsf{T}}P + P\tilde{A})\zeta= -\zeta^{\mathsf{T}}Q\zeta. Integrating along a trajectory that stays in Ωad\Omega_{\mathrm{ad}} on [0,t][0,t] gives ∫0tL(y(s;x)) ds=γqex(x)−γqex(y(t;x))\int_0^t L(y(s;x))\,\mathrm{d}s = \gamma q_{\mathrm{ex}}(x) - \gamma q_{\mathrm{ex}}(y(t;x)). Let the trajectory of xx stay in Ωad\Omega_{\mathrm{ad}} for all t≥0t \ge0. Substituting in (6) with the shaping cost, V(x)=γqex(x)+sup⁡t≥0g(t)V(x) = \gamma q_{\mathrm{ex}}(x) + \sup_{t\ge0}g(t) with g(t)=Ct(x)+(h−γqex)(y(t;x))g(t) = C_t(x) + (h - \gamma q_{\mathrm{ex}})(y(t;x)). By Theorem 2, dist⁡(y(t;x),M)→0\operatorname{dist}(y(t;x),M) \to0 and z(y(t;x))→0z(y(t;x)) \to0. Since hh vanishes on a neighbourhood of the compact set MM and ρ\rho vanishes on {γqex<cI/2}\{\gamma q_{\mathrm{ex}} < c_I/2\}, h(y(t;x))=0h(y(t;x)) = 0 and ρ(y(t;x))=0\rho(y(t;x)) = 0 for all large tt, so Ct(x)C_t(x) converges to a finite C∞(x)C_\infty(x) and g(t)→C∞(x)g(t) \to C_\infty(x). Replacing h−γqexh - \gamma q_{\mathrm{ex}} by ψ=relu⁡(h−γqex)\psi= \operatorname{relu}(h - \gamma q_{\mathrm{ex}}) changes g(t)g(t) only where h−γqex<0h - \gamma q_{\mathrm{ex}} < 0, and there it gives Ct(x)≤C∞(x)≤sup⁡sg(s)C_t(x) \le C_\infty(x) \le\sup_s g(s). Hence sup⁡tg(t)=sup⁡t[Ct(x)+ψ(y(t;x))]=Φ(x)\sup_t g(t) = \sup_t[C_t(x) + \psi(y(t;x))] = \Phi(x), which is (7) for κ=0\kappa= 0. If the trajectory leaves Ωad\Omega_{\mathrm{ad}}, both sides are infinite by definition. Since Ct≥0C_t \ge0 and ψ≥0\psi\ge0, Φ≥ψ≥0\Phi\ge\psi\ge0 and V≥γqexV \ge\gamma q_{\mathrm{ex}}. For (8) and (15), split the supremum over [0,τ][0,\tau] and [τ,∞)[\tau,\infty) and use y(t;x)=y(t−τ;y(τ;x))y(t;x) = y(t-\tau;y(\tau;x)) and Ct(x)=Cτ(x)+Ct−τ(y(τ;x))C_t(x) = C_\tau(x) + C_{t-\tau}(y(\tau;x)) for t≥τt \ge\tau. In normalized time ss, dt=ds/a\mathrm{d}t = \mathrm{d}s/a, so the same identities hold along the orbit of fˉ\bar{f} with Cˉs\bar{C}_s in place of CtC_t. The convergence part of the proof of Theorem 2 uses only Assumption 1, and its finiteness part uses only the first identity of this lemma, so the argument is not circular. □\square

Substituting (7) in the obstacle equation of Theorem 4.4 of Grüne and Zidani [24] gives, formally, the variational inequality min⁡{−LfclΦ−κρ, Φ−ψ}=0\min\{-L_{f_{\mathrm{cl}}}\Phi-\kappa\rho,\ \Phi-\psi\}=0 on Ωad\Omega_{\mathrm{ad}}. Identity (8) is the dynamic programming principle (4.11) of Grüne and Zidani [24] written for Φ\Phi. While Φ\Phi is only nonincreasing along trajectories in Ωad\Omega_{\mathrm{ad}}, VV keeps the strict decrease of Remark 4.3 of Grüne and Zidani (2015): for every x∈D(M)x \in\mathcal{D}(M) and t≥0t \ge0, V(y(t;x))≤V(x)−γ(qex(x)−qex(y(t;x)))V(y(t;x)) \le V(x) - \gamma(q_{\mathrm{ex}}(x) - q_{\mathrm{ex}}(y(t;x))), hence V(y(t;x))<V(x)V(y(t;x)) < V(x) when x∉Mx \notin M and t>0t > 0.

Proof of Theorem 2. Let the trajectory of x∈Ωadx \in\Omega_{\mathrm{ad}} stay in Ωad\Omega_{\mathrm{ad}} while it exists. By Assumption 1(a), fclf_{\mathrm{cl}} is bounded on the compact set Ω‾ad\overline{\Omega}_{\mathrm{ad}}, so the solution exists for all t≥0t \ge0. By Assumption 1(b), z(y(t;x))=eAtz(x)→0z(y(t;x)) = e^{At}z(x) \to0. If dist⁡(y(tj;x),M)\operatorname{dist}(y(t_j;x),M) stays bounded away from zero along a sequence tj→∞t_j \to\infty, a subsequence of y(tj;x)y(t_j;x) converges to a point of Ω‾ad\overline{\Omega}_{\mathrm{ad}} with z=0z = 0, that is, a point of MM. This point would lie at positive distance from MM, which is impossible. Hence dist⁡(y(t;x),M)→0\operatorname{dist}(y(t;x),M) \to0 and x∈D(M)x \in\mathcal{D}(M). This argument uses no property of the internal dynamics. The converse holds by the definition of D(M)\mathcal{D}(M). A forward invariant subset of Ωad\Omega_{\mathrm{ad}} consists of states whose trajectories never leave Ωad\Omega_{\mathrm{ad}}, which gives the maximality. Let x∈D(M)x \in\mathcal{D}(M). The closure of its trajectory consists of the trajectory and its limit points, which lie in M⊂ΩadM \subset\Omega_{\mathrm{ad}} by Assumption 1(c). It is therefore a compact subset of Ωad\Omega_{\mathrm{ad}}, and the continuous function hh is bounded on it. By the first identity of Lemma 1, ∫0tL(y(s;x)) ds≤γqex(x)\int_0^t L(y(s;x))\,ds \le\gamma q_{\mathrm{ex}}(x) for all tt. The shaping cost vanishes on {γqex<cI/2}\{\gamma q_{\mathrm{ex}} < c_I/2\}, which the trajectory enters and does not leave because qex(y(t;x))q_{\mathrm{ex}}(y(t;x)) is nonincreasing and tends to zero, so its integral is finite. Hence V(x)<∞V(x) < \infty. Conversely, V(x)<∞V(x) < \infty implies by definition that the trajectory of xx stays in Ωad\Omega_{\mathrm{ad}}.

Proof of Theorem 3. (i) Let x∈Ix \in\mathcal{I}. By hypothesis, xx satisfies the regularity constraints, so x∈Ωadx \in\Omega_{\mathrm{ad}}. Let t∗=sup⁡{t≥0:y(s;x)∈Ωad for all s∈[0,t]}t_* = \sup\{t \ge0 : y(s;x) \in\Omega_{\mathrm{ad}}\text{ for all }s \in[0,t]\}. By Lemma 1, qexq_{\mathrm{ex}} is nonincreasing along the trajectory on [0,t∗)[0,t_*), so γqex(y(t;x))≤γqex(x)<cI\gamma q_{\mathrm{ex}}(y(t;x)) \le\gamma q_{\mathrm{ex}}(x) < c_I there. Suppose t∗<∞t_* < \infty. Since Ωad\Omega_{\mathrm{ad}} is compact and fclf_{\mathrm{cl}} is smooth on a neighbourhood of it, y(t∗;x)y(t_*;x) exists, lies in ∂Ωad⊂B∩XU\partial\Omega_{\mathrm{ad}} \subset B \cap X_U, and has γqex<cI\gamma q_{\mathrm{ex}} < c_I by continuity. By hypothesis it satisfies the regularity constraints, so it lies on a nonperiodic face of BB, and fclf_{\mathrm{cl}} points strictly into BB across every active face there. Then the trajectory lies outside BB just before t∗t_*, which contradicts y(t;x)∈B∘y(t;x) \in B^\circ for t<t∗t < t_*. Hence t∗=∞t_* = \infty, the trajectory stays in Ωad∩{γqex<cI}⊂I\Omega_{\mathrm{ad}} \cap\{\gamma q_{\mathrm{ex}} < c_I\} \subset\mathcal{I}, so I\mathcal{I} is forward invariant, and I⊂D(M)\mathcal{I} \subset\mathcal{D}(M) by Theorem 2. The points of MM have qex=0q_{\mathrm{ex}} = 0 and lie in Ωad⊂B∘∩XU\Omega_{\mathrm{ad}} \subset B^\circ\cap X_U, so M⊂IM \subset\mathcal{I}. (ii) The inequality v~<c~1\tilde v < \tilde c_1 means −ln⁡(1−v~)<cI-\ln(1-\tilde v) < c_I, and γqex≤−ln⁡(1−v~)\gamma q_{\mathrm{ex}} \le-\ln(1-\tilde v). (iii) Let x∈Ωadx \in\Omega_{\mathrm{ad}} with v~(x)<c2\tilde v(x) < c_2, and define states x(0)=x,x(1),…x^{(0)} = x, x^{(1)}, \ldots as follows. If x(i)∈Ix^{(i)} \in\mathcal{I} or v~(x(i))<c1\tilde v(x^{(i)}) < c_1, which implies x(i)∈Ix^{(i)} \in\mathcal{I} by (ii), stop. Otherwise x(i)∈Ωad∖Ix^{(i)} \in\Omega_{\mathrm{ad}} \setminus\mathcal{I} and c1≤v~(x(i))<c2c_1 \le\tilde v(x^{(i)}) < c_2, so for some Hi∈{1,2,4,8}H_i \in\{1,2,4,8\} the orbit of fˉ\bar f from x(i)x^{(i)} stays in Ωad\Omega_{\mathrm{ad}} over [0,Hi][0,H_i] and x(i+1)=yˉ(Hi;x(i))x^{(i+1)} = \bar y(H_i;x^{(i)}) has v~(x(i+1))<v~(x(i))<c2\tilde v(x^{(i+1)}) < \tilde v(x^{(i)}) < c_2. If the construction stops at some x(i)∈Ix^{(i)} \in\mathcal{I}, the orbit stays in Ωad\Omega_{\mathrm{ad}} afterwards because I\mathcal{I} is forward invariant by (i). Otherwise the steps, each of length at least one, cover all normalized time. In both cases the orbit of fˉ\bar f from xx stays in Ωad\Omega_{\mathrm{ad}} for all normalized time. By Assumption 1(a), aa is bounded on Ωad\Omega_{\mathrm{ad}} by some amax⁡a_{\max}, so the physical time t(s)=∫0sa(yˉ(s′;x))−1 ds′≥s/amax⁡t(s) = \int_0^s a(\bar y(s';x))^{-1}\,ds' \ge s/a_{\max} diverges, the trajectory of fclf_{\mathrm{cl}} stays in Ωad\Omega_{\mathrm{ad}} for all t≥0t \ge0, and x∈D(M)x \in\mathcal{D}(M) by Theorem 2.

Proposition 4 Fix a measurable function v~:B→[0,1]\tilde v : B \to[0,1], levels 0<c1≤c~2<10 < c_1 \le\tilde c_2 < 1 and a measurable counterexample test before drawing NN i.i.d. samples from the uniform distribution P\mathbb{P} on BB. Let c2c_2 be the minimum of c~2\tilde c_2 and the values of v~\tilde v at the sampled counterexamples with v~≥c1\tilde v \ge c_1, and let CC be the set of counterexamples of the test, as implemented, in {c1≤v~<c2}\{c_1 \le\tilde v < c_2\}. Then PN[P(C)>ε]≤(1−ε)N\mathbb{P}^N[\mathbb{P}(C) > \varepsilon] \le(1-\varepsilon)^N for every ε∈(0,1)\varepsilon\in(0,1).

Proof of Proposition 4. Let G(c)=P(counterexamples in {c1≤v~<c})G(c) = \mathbb{P}(\text{counterexamples in }\{c_1 \le\tilde v < c\}), nondecreasing and left-continuous. If G(c~2)≤εG(\tilde c_2) \le\varepsilon there is nothing to prove. Otherwise let c∗=inf⁡{c:G(c)>ε}c_* = \inf\{c : G(c) > \varepsilon\}. The set C∗C_* of counterexamples in {c1≤v~≤c∗}\{c_1 \le\tilde v \le c_*\} has measure at least ε\varepsilon, the event $G(c_2) > \varepsilon requires c2>c∗c_2 > c_* and hence no sample in C∗C_*, and that has probability at most (1−ε)N(1-\varepsilon)^N. This is the one-dimensional case of the scenario theorem of Campi and Garatti [19]. □

Proposition 5 Let the hypotheses of Theorem 3(i) hold, and fix a measurable set E⊂B\mathcal{E} \subset B and a horizon tmax⁡>0t_{\max} > 0 in normalized time before drawing NN i.i.d. samples from P\mathbb{P}. Let

Eu=E∖{x∈Ωad:∃t∈[0,tmax⁡] with yˉ(t;x)∈I and yˉ(s;x)∈Ωad for all s∈[0,t]}.\mathcal{E}_{u}=\mathcal{E}\setminus\{x\in\Omega_{\mathrm{ad}}:\exists t\in[0,t_{\max}]\text{ with }\bar{y}(t;x)\in\mathcal{I}\text{ and }\bar{y}(s;x)\in\Omega_{\mathrm{ad}}\text{ for all }s\in[0,t]\}.

Then E∖D(M)⊂Eu\mathcal{E}\setminus\mathcal{D}(M)\subset\mathcal{E}_{u}. Let KK be the number of samples in Eu\mathcal{E}_{u}, and for j<Nj<N let εj\varepsilon_j solve Pr⁡[Bin⁡(N,εj)≤j]=10−6\Pr[\operatorname{Bin}(N,\varepsilon_j)\le j]=10^{-6}, with εN=1\varepsilon_N=1. Then PN[P(Eu)>εK]≤10−6\mathbb{P}^N[\mathbb{P}(\mathcal{E}_{u})>\varepsilon_K]\le10^{-6}.

Proof of Proposition 5. A state of E∖Eu\mathcal{E}\setminus\mathcal{E}_{u} lies in Ωad\Omega_{\mathrm{ad}} and reaches I\mathcal{I} along a trajectory in Ωad\Omega_{\mathrm{ad}}. By Theorem 3(i), I\mathcal{I} is forward invariant and lies in D(M)\mathcal{D}(M), so the state lies in D(M)\mathcal{D}(M), and E∖D(M)⊂Eu\mathcal{E}\setminus\mathcal{D}(M)\subset\mathcal{E}_{u}. The count KK of samples in Eu\mathcal{E}_{u} is binomial with parameters NN and εˉ=P(Eu)\bar{\varepsilon}=\mathbb{P}(\mathcal{E}_{u}). Since εj\varepsilon_j increases with jj, the event εˉ>εK\bar{\varepsilon}>\varepsilon_K is the event K≤j∗K\le j_*, with j∗j_* the largest jj such that εj<εˉ\varepsilon_j<\bar{\varepsilon}, and it is empty if there is no such jj. The distribution function of the binomial at j∗j_* decreases in its success probability, so Pr⁡[K≤j∗]≤Pr⁡[Bin⁡(N,εj∗)≤j∗]=10−6\Pr[K\le j_*]\le\Pr[\operatorname{Bin}(N,\varepsilon_{j_*})\le j_*]=10^{-6}. This is the one-sided Clopper–Pearson bound. □

Baselines and Closed Loops with an Equilibrium

Zubov PINN. The Zubov PINN follows Liu et al. [34]: two hidden layers of 30 tanh units, W=tanh⁡(σV)W=\tanh(\sigma V), residual ∇W⋅f+σω(1−W2)=0\nabla W\cdot f+\sigma\omega(1-W^2)=0, a boundary term, and a data term on 3000 simulated states. It is trained for 2×1042\times10^4 Adam steps on batches of 2048 from 10510^5 collocation points. On the TFL closed loops the residual is divided by a(x)a(x) of § 2, and the running cost is ω=ζ⊤Qζ\omega=\zeta^\top Q\zeta. The inputs are the normalized state with angles as sine and cosine, and four snapshots are screened. Both PINNs are set to one outside Ωad\Omega_{\mathrm{ad}}.

Dirichlet-Zubov PINN. We follow Meng and Liu [37]. With dmin⁡=min⁡idi/median⁡idid_{\min}=\min_i d_i/\operatorname{median}_i d_i and their scaling constant 0.1, the same residual is solved with the field 0.1dmin⁡f0.1d_{\min}f. The function is zero at the target and one on the 2000 admissible states closest to the boundary of Ωad\Omega_{\mathrm{ad}}. The data term integrates ω/(0.1dmin⁡)\omega/(0.1d_{\min}) along the simulated trajectories, collocation and the boundary term act on admissible states, and the residual is divided by 1+∥0.1dmin⁡Sf∥2+(σω)2\sqrt{1+\lVert0.1d_{\min}S f\rVert^2+(\sigma\omega)^2}, a positive weight.

Scale search. Both PINNs use σ=3/(ςq90)\sigma=3/(\varsigma q_{90}), with q90q_{90} as in the definition of γ\gamma. The divisor ς\varsigma is searched on seed 0 by the coverage after level selection and the final test, with the scoring of the reported numbers. For the Zubov PINN the grid is {1,3,10,30,100,300}\{1,3,10,30,100,300\}. For the Dirichlet-Zubov PINN it is {1,3,10,30,100}\{1,3,10,30,100\}, extended to 300 when 100 is best and to 1000 when 300 is best. Ties go to the smallest ς\varsigma. On the circle and the quadrotor the Zubov PINN ties at the inner layer at every scale, as does the Dirichlet-Zubov PINN on the sinusoid and the quadrotor. On the sinusoid, the smaller divisor ς=0.3\varsigma=0.3 gives no gain for the Zubov PINN.

Larger PINNs and the oracle level. On the TFL closed loops we also trained both PINNs on seed 0 with the features and the network of the candidate, at ς=1\varsigma=1 and ς=10\varsigma=10. The best of these four runs covers 18.2%, 2.5% and 0.0% on the three cases. To bound what any choice of level could give the PINNs there, we also compute an oracle level for every training snapshot of every PINN run. These runs are those of the scale searches, the four seeds and the larger networks. It is the largest level at which no reference state outside the simulated region of attraction lies in the estimate, so it uses the reference labels that the method never sees. The best of these estimates covers 20.2%, 33.7% and 2.4% on the three cases.

Other baselines on the TFL closed loops. The no-decomposition baseline keeps the features of the candidate and its factor tanh⁡(qex/q0)\tanh(q_{\mathrm{ex}}/q_0). Its function is Vθ=max⁡{h,Ψθ}V_\theta= \max\{h,\Psi_\theta\}, and the running cost L\mathcal{L} is integrated along the steps and included in the target. Like sum-of-squares (Xue et al. [53]), moment methods (Korda et al. [30]) require polynomial dynamics and are not applied to the TFL closed loops (§ 1).

Scoring. We score our reimplementations of both PINNs with the level selection and the final test of this paper. For the PINNs, the no-decomposition baseline and the safety-value baseline, c1c_1 is set on samples, because their functions have no lower bound γqex\gamma q_{\mathrm{ex}}. The PINNs are not clipped at zero, because clipping creates plateaus on which no step can decrease, and the oracle level above bounds what any choice of level could give them.

Closed loops with an equilibrium. Algorithm 1 also applies to a closed loop x˙=f(x)\dot{x}=f(x) with an exponentially stable equilibrium at the origin. Its admissible set Ωad\Omega_{\mathrm{ad}} consists of the states of B∘B^\circ that satisfy the state constraints and avoid the obstacles, and hh is built as in (5). We take q(x)=x~TPx~q(x)=\tilde{x}^{\mathsf{T}}P\tilde{x} with x~=Sx\tilde{x}=Sx, where PP solves A~TP+PA~=−I\tilde{A}^{\mathsf{T}}P+P\tilde{A}=-I for A~=S∂f∂x(0)S−1\tilde{A}=S\frac{\partial f}{\partial x}(0)S^{-1}, and the running cost L=γ∥x~∥2+κρ\mathcal{L}=\gamma\lVert\tilde{x}\rVert^2+\kappa\rho, where γ\gamma is that of the step setting, with q90q_{90} the 90th percentile of qq over admissible samples. Since Lfq≠−∥x~∥L_fq\ne-\lVert\tilde{x}\rVert away from the origin, the integral of L\mathcal{L} is no longer known in closed form. The candidate (9) and the target with an accumulated cost (Appendix § 12) therefore use qq in place of qexq_{\mathrm{ex}} and Cs=∫0sL dt+γ(q(y(s))−q(x))C_s=\int_0^s\mathcal{L}\,dt+\gamma(q(y(s))-q(x)), and x~\tilde{x} replaces ζ\zeta in the features. For the same reason, CsC_s can decrease along a trajectory, so the fixed point of the target can overestimate Φ\Phi, for every step (Appendix § 12). The network, the training and the fine-tuning are otherwise unchanged. On these systems the inner layer is a sublevel set of qq at a level set on samples, so it is not verified. Moreover, Theorem 2 is not available, because a trajectory that stays in Ωad\Omega_{\mathrm{ad}} need not converge. The estimates are therefore tested on samples only.

Systems. Each system has its equilibrium at the origin. Leaving the box counts as leaving Ωad\Omega_{\mathrm{ad}}. The reversed Van der Pol oscillator is x˙1=−x2\dot{x}_1=-x_2, x˙2=x1−μ(1−x12)x2\dot{x}_2=x_1-\mu(1-x_1^2)x_2, on [−2.5,2.5]×[−3.5,3.5][-2.5,2.5]\times[-3.5,3.5] for μ=1\mu=1 and [−3,3]×[−6,6][-3,3]\times[-6,6] for μ=3\mu=3. The pendulum is x˙1=x2\dot{x}_1=x_2, x˙2=sin⁡x1−x2+u\dot{x}_2=\sin x_1-x_2+u, with u=sat⁡0.8(−4.4142x1−2.3163x2)u=\operatorname{sat}_{0.8}(-4.4142x_1-2.3163x_2), the gains of Example 5 of Liu et al. [34] and a torque bound that we add, on [−π,π]×[−4,4][-\pi,\pi]\times[-4,4]. The 3-D system is x˙1=−x1+x2\dot{x}_1=-x_1+x_2, x˙2=0.1x1−2x2−x12−0.1x13\dot{x}_2=0.1x_1-2x_2-x_1^2-0.1x_1^3, x˙3=−x3+x1x2\dot{x}_3=-x_3+x_1x_2, with x3<1x_3<1 and (x1+1)2+(x2−1)2>0.32(x_1+1)^2+(x_2-1)^2>0.3^2, on [−3,3]2×[−2,2][-3,3]^2\times[-2,2]. The power system is x˙1=x2\dot{x}_1=x_2, x˙2=−0.5x2−sin⁡(x1+π/3)+sin⁡(π/3)\dot{x}_2=-0.5x_2-\sin(x_1+\pi/3)+\sin(\pi/3), on [−2,3]×[−3,3][-2,3]\times[-3,3]. The reference simulation integrates 5×1045\times10^4 states with RK4, step 0.010.01 and horizon 4040 (60 for the power system), and a state has converged when the scaled state has norm below 10−310^{-3} for one time unit.

Results. We ran Algorithm 1 on seed 0 of the five systems (Table 3). Its step setting keeps the long step on two of them and the short step on the others. On every system, as on the TFL cases, the share of box samples ranks the two settings as coverage does. After fine-tuning, the estimates of the kept setting cover between 86.5%86.5\% and 98.9%98.9\% of the simulated regions of attraction, and no reference state lies inside one of them and outside the simulated region of attraction. We ran one seed per system and did not compare with baselines on these systems.

Interval Verification of the Inner Layer

The hypotheses of Theorem 3(i) are statements about explicit functions on a box, and we check them by interval branch and bound (Moore et al. [41]). The box is split into cells. A cell is discarded when interval arithmetic shows that γqex≥cI\gamma q_{\mathrm{ex}} \ge c_I on it or that it contains no violating point. A cell contains no violating point when all distances did_i to the regularity constraints are positive on it and, if it touches a face, the normal component of the field has the inward sign. The remaining cells are bisected along their widest normalized side until none is left. The checks run on the closed box with the periodic angle over [−π,π][-\pi,\pi]. The lower bound of the quadratic form is the larger of two bounds: the one from the eigendecomposition of PP, with each squared projection of ζ\zeta bounded below by zero, and λmin⁡(P)∑iζi2\lambda_{\min}(P)\sum_i\zeta_i^2. The distances did_i and the field components are the symbolic expressions of Appendix § 8, evaluated in float64 interval arithmetic with outward rounding by two units in the last place per operation and with enclosed constants. For each case the level is found by bisection over twelve rounds between zero and the sampled minimum of γqex\gamma q_{\mathrm{ex}} over samples outside Ωad\Omega_{\mathrm{ad}} and samples on box faces where the field points outward. The last verified level is multiplied by 0.99. This one percent reduction absorbs the float32 evaluation of qexq_{\mathrm{ex}} in the training pipeline and the residual effects of the eigendecomposition of PP and the float32 storage of PP and SzS_z. Table 4 lists the verified levels, the share of the box in I\mathcal{I}, the number of cells and the time. The verified claim concerns the set {qex<cI/γ}\{q_{\mathrm{ex}} < c_I/\gamma\} and does not otherwise involve γ\gamma. The verification was run with the cost scale 3/q903/q_{90} of the short-step setting, and for the long-step setting the level cIc_I is one third of the verified one, which describes the same set. Table 4 gives the levels for γ=1/q90\gamma= 1/q_{90}.

CasecIc_Ic1c_1I\mathcal{I} (% box)CellsTime (s)
Circle0.03500.03445.40944395588
Sinusoid0.01700.01681.011047557110
Quadrotor0.00020.00020.00212709813626

Table 4. Verified inner layers of the TFL cases, with the cells and time of the final verification.

Configuration

Training setup. The candidate uses the network, optimizer and collocation points of Table 5, with the level selection and fine-tuning of § 3. We set the configuration on seed 0, with level selection and a final test, while watching the coverage of the reference simulation. In a first stage we chose the weight of the finite-step penalty and Vmax⁡V_{\max}, which gave the short-step setting. In a second stage, with these weights fixed, we chose the step τ\tau, the cost scale γ\gamma and the exit weight wexw_{\mathrm{ex}} one at a time on the quadrotor, which gave the long-step setting. The other constants of Table 5 and Table 6 come from an earlier version of the method on the unicycle. All reported runs, seed 0 included, were trained after these choices. Since the reference simulation guided these choices on seed 0, the coverage of seed 0 is not an independent test of the configuration, and Table 7 lists every seed.

Value function and candidate
Value functionbi=0.1⋅median⁡dib_i = 0.1\cdot\operatorname{median} d_i, h=∑irelu⁡(bi/di−1)2h = \sum_i \operatorname{relu}(b_i/d_i - 1)^2, γ\gamma of the step setting, with q90q_{90} the 90th percentile of qexq_{\mathrm{ex}} over admissible samples
Shaping costκ=0.3\kappa= 0.3 if the inner layer holds at least 1% of 2⋅1052\cdot10^5 box samples, and κ=0\kappa= 0 otherwise. The function ρ\rho is a smooth step from γqex=cI/2\gamma q_{\mathrm{ex}} = c_I/2 to cIc_I. The share on the sinusoid is 1.01%, close to this threshold
NetworkNθ\mathcal{N}_{\theta}: 64-64-64 tanh. Features: normalized state (angles as sine and cosine), tanh⁡(ζ/3)\tanh(\zeta/3), tanh⁡(12ln⁡(di/median⁡di))\tanh\left(\frac{1}{2}\ln(d_i/\operatorname{median} d_i)\right) for the constraints on ΔD\Delta D and ∥∇ϖ∥−1\lVert\nabla\varpi\rVert^{-1}. Softplus sharpness β=10\beta= 10. The constant q0q_0 is the first percentile of qexq_{\mathrm{ex}} on 2⋅1052\cdot10^5 collocation points
InitializationThe network starts close to Φθ=ψ\Phi_{\theta} = \psi, which lies below the exact solution
Training
Collocation4⋅1054\cdot10^5 Sobol box samples, 10510^5 near the faces, 10510^5 next to the singular set (admissible ones kept)
Step settingLong: τ=2\tau= 2 (80 RK4 substeps), γ=1/q90\gamma= 1/q_{90}, wex=10w_{\mathrm{ex}} = 10. Short: τ=0.2\tau= 0.2 (8 RK4 substeps), γ=3/q90\gamma= 3/q_{90}, wex=0w_{\mathrm{ex}} = 0. Each system keeps the setting whose seed-0 estimate, after level selection and the final test and before fine-tuning, holds the larger share of box samples. The regression loss and the obstacle penalty of a point whose step leaves Ωad\Omega_{\mathrm{ad}} are multiplied by 1+wex1+w_{\mathrm{ex}}
Target networkCopy of the averaged weights (decay 0.999), refreshed every epoch
OptimizationAdam, 100 epochs, batch 4096, cosine learning rate 10−3→10−510^{-3}\to10^{-5}. Exponential loss weight e2min⁡(Vθ,Vmax⁡)e^{2\min(V_{\theta},V_{\max})} with Vmax⁡=3V_{\max}=3 (no such weight on the quadrotor)
Decrease penaltyWeight 10, on relu⁡((LfclΦθ+κρ)/a)e−Vθ\operatorname{relu}((L_{f_{\mathrm{cl}}}\Phi_{\theta}+\kappa\rho)/a)e^{-V_{\theta}}, where Ψθ>ψ\Psi_{\theta}>\psi
Obstacle penaltyWeight 10, on relu⁡(ψ+0.01−Ψθ)e−Vθ\operatorname{relu}(\psi+0.01-\Psi_{\theta})e^{-V_{\theta}}, where Ψθ≤ψ\Psi_{\theta}\le\psi and Lfclψ+κρ>0L_{f_{\mathrm{cl}}}\psi+\kappa\rho>0
Finite-step penaltyWeight 100 (10 on the quadrotor), unit step with 40 RK4 substeps, on relu⁡(e−(Vθ(x)−χ[γrelu⁡(qex(x)−qex(x+))+C1])−e−Vθ(x+))\operatorname{relu}\left(e^{-(V_{\theta}(x)-\chi[\gamma\operatorname{relu}(q_{\mathrm{ex}}(x)-q_{\mathrm{ex}}(x^+))+C_1])}-e^{-V_{\theta}(x^+)}\right) with χ=0.2\chi=0.2 and x+x^+ the end of the unit step from xx, on steps in Ωad\Omega_{\mathrm{ad}} with Vθ(x)<8V_{\theta}(x)<8
Violating statesEvery 5 epochs up to epoch 90: the 10410^4 states whose unit step violates the decrease most (no margin), in a pool of 2⋅1052\cdot10^5 box samples and 5⋅1045\cdot10^4 next to the singular set, 4 Gaussian neighbours (standard deviation 0.01 of the half-widths), rolling buffer 10510^5

Table 5. Configuration of every case: value function, candidate and training. Step lengths are in normalized time.

Level selection and fine-tuning
Level selectionHorizons H∈{1,2,4,8}H \in\{1,2,4,8\}, RK4 substep 0.025, states with ∥ζ∥≤10−3\lVert\zeta\rVert\le10^{-3}, which lie in I\mathcal{I}, exempt from the decrease. Snapshots every 5 epochs and at epoch 0. A test on a number of box samples adds half as many near the faces, a fifth as many next to the singular set and half as many on the faces, where states with outward field count as counterexamples. Screening: 5⋅1045 \cdot10^{4} box samples, 3 rounds. Retest of the top 3: 10510^{5} box samples, 5 rounds. Ranking by share of 5⋅1045 \cdot10^{4} fixed box samples. Final test: 10610^{6} i.i.d. uniform box samples
Fine-tuning4 rounds, each on the counterexamples with c1≤vθ<min⁡{c2+0.05,0.999}c_{1} \le v_{\theta} < \min\{c_{2}+0.05,0.999\} for a level c2c_{2} estimated again in that round, pool of 10610^{6} box samples, 5⋅1055 \cdot10^{5} near the faces and 2⋅1052 \cdot10^{5} next to the singular set, 8 Gaussian neighbours per counterexample, 300 Adam steps (learning rate 5⋅10−55 \cdot10^{-5}, batch 4096), weight 100 on counterexamples whose step stays in Ωad\Omega_{\mathrm{ad}}, weight 10 on those that leave it, decrease margin 0.2, counterexample sets capped at 5⋅1045 \cdot10^{4}, anchor on 2⋅1052 \cdot10^{5} box samples, exponential loss weight with Vmax⁡=3V_{\max}=3 on all systems
Reference simulation
Reference5⋅1045 \cdot10^{4} Sobol states, RK4 step 0.01, horizon 30 (unicycle) and 40 (quadrotor). Converged: held for one time unit, ∥ξ∥<10−3\lVert\xi\rVert< 10^{-3} and ∣η2−1∣<0.05\lvert\eta_{2}-1\rvert< 0.05 (unicycle), ∥ζ∥<10−3\lVert\zeta\rVert< 10^{-3} (quadrotor). States that have neither converged nor failed at the horizon count as outside

Table 6. Configuration of every case: level selection, fine-tuning and reference simulation. Step lengths are in normalized time, except in the reference simulation.

CaseSeedc2c_2 beforeCov. beforeRoundc2c_2 sel.c2c_2 finalAnnulusCov.FP
Circle00.869571.9940.97000.970057452074.040.000
Circle10.857069.1640.97680.976854482670.460.000
Circle20.853969.8230.97860.973957055272.520.000
Circle30.835869.2340.96870.968754819970.850.000
Sinusoid00.907182.5310.97700.977078482383.750.000
Sinusoid10.866881.6410.97220.970079075184.010.000
Sinusoid20.877780.9620.96090.960977292582.310.000
Sinusoid30.909983.1640.96970.967778767283.650.000
Quadrotor00.314024.8600.39220.250615173122.860.000
Quadrotor10.122613.5300.29910.122613586813.530.000
Quadrotor20.343123.7300.36890.247213656721.030.000
Quadrotor30.254123.1500.34720.182314581119.070.000

Table 7. Per-seed results. Before: the trained network. Round: selected fine-tuning round (0 is the trained network). Sel. and Final: level after the retest and after the final test. Annulus: final-test samples between c1c_1 and the selected level. FP as in Table 1.

Step setting. Unrolling the target over successive steps, with the shaping cost below, shows that its only continuous fixed point that vanishes on MM is sup⁡t≥0[Ct(x)+ψ(y(t;x))]\sup_{t\geq0}[C_t(x)+\psi(y(t;x))] for every step τ\tau. The candidate satisfies both conditions, because Ψθ\Psi_{\theta} and ψ\psi vanish on MM. On a TFL closed loop this function is the safety value function Φ=V−γqex\Phi=V-\gamma q_{\mathrm{ex}}. Indeed, qexq_{\mathrm{ex}} is the exact cost-to-go of the error dynamics, so the cost accumulated beyond γqex\gamma q_{\mathrm{ex}} is the nonnegative shaping cost alone. Hence CtC_t is nondecreasing in tt, and the margin function ψ=relu⁡(h−γqex)\psi=\operatorname{relu}(h-\gamma q_{\mathrm{ex}}) can replace h−γqexh-\gamma q_{\mathrm{ex}} in the supremum (proof of Lemma 1). A long step then reaches the fixed point in fewer epochs, because each update carries the value of the end point over a long arc of the orbit. When the quadratic form is only a local Lyapunov function, as for the closed loops with an equilibrium of Appendix D, CtC_t can decrease along a trajectory. The fixed point then overestimates Φ\Phi for every step, so no step is exact there. In every case the method compares the two settings of § 3.3 on seed 0 by the share of box samples in the estimate, after level selection and the final test and before fine-tuning. It then keeps the setting with the larger share for all seeds. The comparison uses no reference labels, and both estimates have passed the same tests. The two settings differ in the cost scale and the exit weight as well as in the step, because each was developed as a whole (see the training setup above). Table 3 in Appendix D lists the comparison for every system.

Shaping cost. When the inner layer holds at least one percent of the box, we add a shaping cost κρ\kappa\rho with κ=0.3\kappa=0.3 to L\mathcal{L}, to steepen vv just outside the inner layer. Otherwise κ=0\kappa=0, as on the quadrotor. The quadrotor also uses the weight 10 for the finite-step penalty, against 100 on the unicycle, and no exponential loss weight during training. Here ρ∈[0,1]\rho\in[0,1] is a smooth nondecreasing function of γqex\gamma q_{\mathrm{ex}}, equal to zero below cI/2c_I/2 and to one above cIc_I, where cIc_I is the level of the inner layer. Its integral is not known in closed form, so Φ\Phi absorbs it. With Ct(x)=κ∫0tρ(y(s;x)) dsC_t(x)=\kappa\int_{0}^{t}\rho(y(s;x))\,\mathrm{d}s in physical time, the decomposition V=γqex+ΦV=\gamma q_{\mathrm{ex}}+\Phi holds with Φ(x)=sup⁡t≥0[Ct(x)+ψ(y(t;x))]\Phi(x)=\sup_{t\ge0}[C_t(x)+\psi(y(t;x))], under the conventions of Lemma 1, and (8) becomes

Φ(x)=max⁡{max⁡0≤s≤τ[Cs(x)+ψ(y(s;x))],Cτ(x)+Φ(y(τ;x))}.(15)\Phi(x)=\max\left\{\max_{0\le s\le\tau}\left[C_s(x)+\psi(y(s;x))\right], C_\tau(x)+\Phi(y(\tau;x))\right\}. \tag*{(15)}

In normalized time the accumulated cost is Cˉs(x)=∫0s(κρ/a)(yˉ(s′;x)) ds′\bar{C}_s(x)=\int_{0}^{s}\left(\kappa\rho/a\right)(\bar{y}(s';x))\,\mathrm{d}s'. Training then stores Cˉτ(x)\bar{C}_\tau(x) with the end point, the running maximum is Rτ(x)=max⁡0≤s≤τ[Cˉs(x)+ψ(yˉ(s;x))]R_\tau(x)=\max_{0\le s\le\tau}[\bar{C}_s(x)+\psi(\bar{y}(s;x))], and the target is U(x)=exp⁡(−max⁡{Rτ(x),Cˉτ(x)+Φθ(xτ)})\mathcal{U}(x)=\exp(-\max\{R_\tau(x),\bar{C}_\tau(x)+\Phi_\theta(x_\tau)\}). The decrease penalty asks for LfcΦθ≤−κρL_{f_c}\Phi_\theta\le-\kappa\rho and the obstacle penalty for Ψθ∘ψ\Psi_\theta\circ\psi where Lfcψ>−κρL_{f_c}\psi>-\kappa\rho. For κ=0\kappa=0 all of these reduce to the expressions of § 3. Theorem 3 is unaffected, because Vθ≥γqexV_\theta\ge\gamma q_{\mathrm{ex}} still holds.

Penalties. Lie derivatives in the decrease penalty and the obstacle penalty are taken at points of differentiability. The finite-step penalty acts on steps of unit length in Ωad\Omega_{\mathrm{ad}} from xx to x+x^{+}. It asks for Vθ(x+)≤Vθ(x)−χ[γ(qex(x)−qex(x+))+C1]V_\theta(x^{+})\le V_\theta(x)-\chi[\gamma(q_{\mathrm{ex}}(x)-q_{\mathrm{ex}}(x^{+}))+C_1] with χ=0.2\chi=0.2 and C1C_1 the accumulated shaping cost over the unit step, and it is applied to the values e−Vθe^{-V_\theta}. An error in VθV_\theta at a level VV moves vθv_\theta by only the factor e−Ve^{-V} of it, so without the exponential loss weight the loss would carry almost no gradient at the levels near one. Every five epochs, the 10410^4 states of a fresh pool of 2.5×1052.5 \times10^5 whose unit step violates the decrease most, without the margin χ\chi, join the data. Each comes with four Gaussian neighbours, and the data are kept in a rolling buffer of 10510^5. The network starts close to Φθ=ψ\Phi_\theta= \psi, which lies below the exact solution. Where Φ=ψ\Phi= \psi, any network with Ψθ≤ψ\Psi_\theta\le\psi represents Φ\Phi exactly, so the network has to approximate Φ\Phi only where Φ>ψ\Phi> \psi, on states whose trajectories later come closer to a constraint. Every term carries the exponential loss weight e2min⁡(Vθ,Vmax⁡)e^{2\min(V_\theta,V_{\max})} with Vmax⁡=3V_{\max} = 3, and none during training on the quadrotor. Table 5 gives every constant.

Screening and retest. Screening uses 3×5⋅1043 \times5 \cdot10^4 box samples and the additional samples of Table 6. The three best snapshots are retested on 5×1055 \times10^5 fresh box samples, and the best one is kept with the lowest level it received.

Additional Results

Unsafe samples and the final test. The direct test of unsafe states draws 10610^6 i.i.d. box samples per case and simulates each sample in the union of the four estimates until it reaches the inner layer. Of the 641 790, 815 296 and 135 702 samples in the unions, 1, 1 and 1 leave the box first, and finer integration confirms all 3. These counts give the bounds of § 5, and each bound holds for every seed. By Theorem 3(iii), the annulus of every seed whose estimate contains an unsafe sample holds counterexamples that the final test missed. The final test found a counterexample and lowered the level in seven of the twelve runs (Table 7).

How reliable is the direct test of unsafe states? It is at least as accurate as the reference simulation. The test uses seeds disjoint from all other tests, the RK4 integration of the normalized field with substeps of length 0.025 as in the final test, and tmax⁡=200t_{\max} = 200. Every sample that reached I\mathcal{I} did so within 8 units of normalized time. On 9000 reference states, the test and the reference simulation disagree on 41, and the reference simulation with a four times finer step agrees with the test on 38 of them. In the other 3, the test counts as unsafe a state that the finer reference finds converging, which is the conservative direction. None of the disagreeing states lies in an estimate, so this comparison does not test the safe labels inside the estimates. The 3 samples counted as unsafe still leave Ωad\Omega_{\mathrm{ad}} with four and sixteen times finer substeps, and the reference simulation with step 0.001 finds that they leave the box.

What do the two tests leave open? With N=106N = 10^6 and ε=ln⁡(106)/N=1.38×10−5\varepsilon= \ln(10^6)/N = 1.38 \times10^{-5}, Proposition 4 says with confidence 1−10−61 - 10^{-6} that the uniform measure of counterexamples in the annulus is at most ε\varepsilon. As noted in § 4, this alone does not bound the measure of states of the outer layer outside D(M)\mathcal{D}(M). The reason is that a function can decrease everywhere except on a set of small measure while most trajectories diverge [29]. Both tests integrate numerically and check membership in Ωad\Omega_{\mathrm{ad}} at substeps of length 0.025, so a crossing of the singular set between substeps could go unnoticed. The test of Proposition 5 also checks the sign of det⁡D\det D at every substep, which detects a sign change of det⁡D\det D between substeps but not a crossing of the other constraints. The union of the outer layers over several seeds is therefore also tested on samples only. A deterministic check of the hypothesis of Theorem 3(iii) needs an enclosure of the flow over a unit step on every cell of the annulus, and we did not compute one.

What limits the outer layer? Both the level and the fit limit it, and the fit matters most on the quadrotor. The exact value function cut at the level of the candidate covers 85.4%, 88.1% and 46.8% of the simulated region of attraction, and the outer layers cover 84%, 95% and 41% of these shares. On the unicycle the levels lie above 0.96, and the states that the exact value function leaves out at these levels have trajectories that pass near a constraint before converging. They therefore have a large Φ\Phi, and only a level closer to one would include them. On the quadrotor, counterexamples hold the level between 0.12 and 0.25, while the exact value function cut at 0.9 covers 80.5%. At these levels the bound Vθ≥γexV_\theta\geq\gamma_{\mathrm{ex}} excludes 37.5% of the simulated region of attraction, against 0.1% on the circle and 0.0% on the sinusoid.

What do the components contribute? Fine-tuning helps on the unicycle and is never selected on the quadrotor, where screening kept the trained network in four of four runs. Each of these networks therefore has two final tests, one before and one after fine-tuning, and its reported level is the lower of the two (Table 7). The draw of the final test alone moves the coverage of a seed by up to 7.0 points. Table 8 removes components on the circle, where the PDE residual loss alone, without fitted value iteration, penalties, counterexamples, exponential loss weight or weight averaging, reaches 10.5 %. The sum form replaces the maximum in (9) by Φθ=ψ+Ψθ\Phi_\theta= \psi+ \Psi_\theta, and it has neither the obstacle penalty nor the one-sided loss of § 3.3, because both refer to the maximum in (9).

Does the integration step change the reference labels? It changes few of them, at most 0.74 % on a case. We simulated 5000 of the 5×1045 \times10^{4} reference states of each case again with an RK4 step of 0.0025 in place of 0.01. The label changes on 0.24 %, 0.24 % and 0.74 % of these states on the circle, the sinusoid and the quadrotor.

What limits the inner layer? Contact with the constraints bounds it from above, and the interval verification stops below that bound. On samples, the sublevel sets of γqex\gamma_{\mathrm{qex}} first meet a regularity constraint, or a face of the box where the field points outward, at levels of at most 0.0417, 0.0188 and 0.0019 on the three cases. The verified levels are 0.0350, 0.0170 and 0.0002 (Table 4), and on the quadrotor the verification at the next level exceeds its budget of cells. The first contact lies near the centre of the circle, on the speed constraint for the sinusoid, and on the zero-thrust face for the quadrotor. Thrust enters zz only through the acceleration coordinates, so a state with no thrust can have a small qexq_{\mathrm{ex}}, and the inner layer of the quadrotor is negligible.

Computation time. All computations run on the CPU of a laptop (Apple M1). The trainings ran eight at a time on one thread each, next to the level selection of other runs, and took 44 to 73 minutes per seed under this load. Level selection took 1 to 13 minutes, fine-tuning 2 to 59 minutes, and the interval verification 110 to 3626 seconds per case.

References

  1. [1]Alessandro Abate, Daniele Ahmed, Alec Edwards, Mirco Giacobbe, and Andrea Peruffo. FOSSIL: A software tool for the formal synthesis of Lyapunov functions and barrier certificates using neural networks. In ACM International Conference on Hybrid Systems: Computation and Control (HSCC), pages 1–11, 2021.
  2. [2]Dirk Aeyels and Joan Peuteman. A new asymptotic stability criterion for nonlinear time-variant differential equations. IEEE Transactions on Automatic Control, 43(7):968–971, 1998.DOI
  3. [4]Adeel Akhtar and Christopher Nielsen. Path following for a car-like robot using transverse feedback linearization and tangential dynamic extension. In IEEE Conference on Decision and Control and European Control Conference (CDC-ECC), pages 7974–7979, 2011.DOI
  4. [5]Adeel Akhtar, Steven L. Waslander, and Christopher Nielsen. Path following for a quadrotor using dynamic extension and transverse feedback linearization. In IEEE Conference on Decision and Control (CDC), pages 3551–3556, 2012.DOI
  5. [6]Adeel Akhtar, Steven L. Waslander, and Christopher Nielsen. Fault tolerant path following for a quadrotor. In IEEE Conference on Decision and Control (CDC), pages 847–852, 2013.DOI
  6. [7]Adeel Akhtar, Christopher Nielsen, and Steven L. Waslander. Path following using dynamic transverse feedback linearization for car-like robots. IEEE Transactions on Robotics, 31(2):269–279, 2015.DOI
  7. [8]Adeel Akhtar, Sajid Saleem, and Jinjun Shan. Path invariant controllers for a quadrotor with a cable-suspended payload using a global parameterization. IEEE Transactions on Control Systems Technology, 30(5):2002–2017, 2022.DOI
  8. [9]Adeel Akhtar, Sajid Saleem, and Jinjun Shan. Path following of a quadrotor with a cable-suspended payload. IEEE Transactions on Industrial Electronics, 70(2):1646–1654, 2023.DOI
  9. [10]Mohamed Al Lawati and Adeel Akhtar. Path-following control of a quadrotor using quasi-static transverse feedback linearization. In European Control Conference (ECC), pages 3454–3459, 2026a.
  10. [11]Mohamed Al Lawati and Adeel Akhtar. Quasi-static fault-tolerant feedback control of a quadrotor under rotor failure with provable safety guarantees. In IEEE Conference on Control Technology and Applications (CCTA), pages 681–686, 2026b.
  11. [12]Somil Bansal and Claire J. Tomlin. DeepReach: A deep learning approach to high-dimensional reachability. In IEEE International Conference on Robotics and Automation (ICRA), pages 1817–1824, 2021.arxiv.org/abs/2011.02082
  12. [13]Matthieu Barreau and Nicola Bastianello. Learning maximal regions of attraction on compact sets using Taylor Lyapunov networks. arXiv preprint arXiv:2408.17246, 2024.arxiv.org/abs/2408.17246
  13. [14]Felix Berkenkamp, Riccardo Moriconi, Angela P. Schoellig, and Andreas Krause. Safe learning of regions of attraction for uncertain, nonlinear systems with Gaussian processes. In IEEE Conference on Decision and Control (CDC), pages 4661–4666, 2016.DOI
  14. [15]Ruxandra Bobiti and Mircea Lazar. Automated-sampling-based stability verification and DOA estimation for nonlinear systems. IEEE Transactions on Automatic Control, 63(11):3659–3674, 2018.DOI
  15. [16]Nicholas M. Boffi, Stephen Tu, Nikolai Matni, Jean-Jacques E. Slotine, and Vikas Sindhwani. Learning stability certificates from data. In Conference on Robot Learning (CoRL), volume 155 of Proceedings of Machine Learning Research, pages 1341–1350, 2021.arxiv.org/abs/2008.05952
  16. [17]Olivier Bokanowski, Averil Prost, and Xavier Warin. Neural networks for first order HJB equations and application to front propagation with obstacle terms. Partial Differential Equations and Applications, 4(5):45, 2023.arxiv.org/abs/2210.04300
  17. [18]Fabio Camilli, Lars Grüne, and Fabian Wirth. A generalization of Zubov’s method to perturbed systems. SIAM Journal on Control and Optimization, 40(2):496–515, 2001.
  18. [19]Marco C. Campi and Simone Garatti. The exact feasibility of randomized solutions of uncertain convex programs. SIAM Journal on Optimization, 19(3):1211–1230, 2008.DOI
  19. [20]Ya-Chien Chang, Nima Roohi, and Sicun Gao. Neural Lyapunov control. In Advances in Neural Information Processing Systems, volume 32, pages 3245–3254, 2019.arxiv.org/abs/2005.00611
  20. [21]Hongkai Dai, Benoit Landry, Lujie Yang, Marco Pavone, and Russ Tedrake. Lyapunov-stable neural-network control. In Robotics: Science and Systems (RSS), 2021.arxiv.org/abs/2109.14152
  21. [22]Timm Faulwasser and Rolf Findeisen. Nonlinear model predictive control for constrained output path following. IEEE Transactions on Automatic Control, 61(4):1026–1039, 2016.DOI
  22. [23]Roman Geiselhart, Rob H. Gielen, Mircea Lazar, and Fabian R. Wirth. An alternative converse Lyapunov theorem for discrete-time systems. Systems & Control Letters, 70:49–59, 2014.DOI
  23. [24]Lars Grüne and Hasnaa Zidani. Zubov’s equation for state-constrained perturbed nonlinear systems. Mathematical Control and Related Fields, 5(1):55–71, 2015.DOI
  24. [26]Kai-Chieh Hsu, Vicenç Rubies-Royo, Claire J. Tomlin, and Jaime F. Fisac. Safety and liveness guarantees through reach-avoid reinforcement learning. In Robotics: Science and Systems (RSS), 2021.DOI
  25. [27]Alberto Isidori. Nonlinear Control Systems. Springer, 3rd edition, 1995.
  26. [28]Wei Kang, Kai Sun, and Liang Xu. Data-driven computational methods for the domain of attraction and Zubov’s equation. IEEE Transactions on Automatic Control, 69(3):1600–1611, 2024.arxiv.org/abs/2112.14415
  27. [29]Joris Kenanian, Ayca Balkan, Raphaël M. Jungers, and Paulo Tabuada. Data driven stability analysis of black-box switched linear systems. Automatica, 109:108533, 2019.arxiv.org/abs/1803.03687
  28. [30]Milan Korda, Didier Henrion, and Colin N. Jones. Inner approximations of the region of attraction for polynomial dynamical systems. In IFAC Symposium on Nonlinear Control Systems (NOLCOS), volume 46 of IFAC Proceedings Volumes, pages 534–539, 2013.arxiv.org/abs/1210.3184
  29. [31]Haoyu Li, Xiangru Zhong, Bin Hu, and Huan Zhang. Two-stage learning of stabilizing neural controllers via Zubov sampling and iterative domain expansion. In Advances in Neural Information Processing Systems, volume 38, pages 82230–82259, 2025.arxiv.org/abs/2506.01356
  30. [32]Jun Liu. Universal approximation of maximal Lyapunov functions with anchored neural networks. arXiv preprint arXiv:2608.17290, 2026.arxiv.org/abs/2608.17290
  31. [33]Jun Liu, Yiming Meng, Maxwell Fitzsimmons, and Ruikun Zhou. LyZNet: A lightweight Python tool for learning and verifying neural Lyapunov functions and regions of attraction. In ACM International Conference on Hybrid Systems: Computation and Control (HSCC), pages 1–8, 2024.
  32. [34]Jun Liu, Yiming Meng, Maxwell Fitzsimmons, and Ruikun Zhou. Physics-informed neural network Lyapunov functions: PDE characterization, learning, and verification. Automatica, 175:112193, 2025.DOI
  33. [35]Ian R. Manchester. Transverse dynamics and regions of stability for nonlinear hybrid limit cycles. In IFAC World Congress, volume 44 of IFAC Proceedings Volumes, pages 6285–6290, 2011.arxiv.org/abs/1010.2241
  34. [36]Yiming Meng and Jun Liu. Lyapunov-barrier characterization of robust reach–avoid–stay specifications for hybrid systems. Nonlinear Analysis: Hybrid Systems, 49:101340, 2023.arxiv.org/abs/2211.00814
  35. [37]Yiming Meng and Jun Liu. Towards learning and verifying maximal Lyapunov-barrier functions with a Zubov PDE formulation. In European Control Conference (ECC), 2026.arxiv.org/abs/2511.09523
  36. [38]Yiming Meng, Yinan Li, Maxwell Fitzsimmons, and Jun Liu. Smooth converse Lyapunov-barrier theorems for asymptotic stability with safety constraints and reach-avoid-stay specifications. Automatica, 144:110478, 2022.arxiv.org/abs/2009.04432
  37. [39]Yiming Meng, Ruikun Zhou, and Jun Liu. Learning regions of attraction in unknown dynamical systems via Zubov–Koopman lifting: Regularities and convergence. IEEE Transactions on Automatic Control, 70(10):6672–6687, 2025.arxiv.org/abs/2311.15119
  38. [40]Ian M. Mitchell, Alexandre M. Bayen, and Claire J. Tomlin. A time-dependent Hamilton–Jacobi formulation of reachable sets for continuous dynamic games. IEEE Transactions on Automatic Control, 50(7):947–957, 2005.DOI
  39. [41]Ramon E. Moore, R. Baker Kearfott, and Michael J. Cloud. Introduction to Interval Analysis. SIAM, 2009.DOI
  40. [42]Christopher Nielsen and Manfredi Maggiore. Maneuver regulation via transverse feedback linearization: Theory and examples. In IFAC Symposium on Nonlinear Control Systems (NOLCOS), volume 37 of IFAC Proceedings Volumes, pages 57–64, 2004.
  41. [43]Christopher Nielsen and Manfredi Maggiore. Output stabilization and maneuver regulation: A geometric approach. Systems & Control Letters, 55(5):418–427, 2006.DOI
  42. [44]Christopher Nielsen and Manfredi Maggiore. On local transverse feedback linearization. SIAM Journal on Control and Optimization, 47(5):2227–2250, 2008.DOI
  43. [45]Meeko Oishi, Ian Mitchell, Claire Tomlin, and Patrick Saint-Pierre. Computing viable sets and reachable sets to design feedback linearizing control laws under saturation. In IEEE Conference on Decision and Control (CDC), pages 3801–3807, 2006.
  44. [46]Spencer M. Richards, Felix Berkenkamp, and Andreas Krause. The Lyapunov neural network: Adaptive stability certification for safe learning of dynamical systems. In Conference on Robot Learning (CoRL), volume 87 of Proceedings of Machine Learning Research, pages 466–476, 2018.arxiv.org/abs/1808.00924
  45. [47]Ashton Roza and Manfredi Maggiore. Path following controller for a quadrotor helicopter. In American Control Conference (ACC), pages 4655–4660, 2012.
  46. [48]Mohamed Serry, Haoyu Li, Ruikun Zhou, Huan Zhang, and Jun Liu. Safe domains of attraction for discrete-time nonlinear systems: Characterization and verifiable neural network estimation. In IEEE Conference on Decision and Control (CDC), pages 5774–5781, 2025.
  47. [49]Mohamed Serry, Maxwell Fitzsimmons, and Jun Liu. Safe and robust domains of attraction for discrete-time systems: A set-based characterization and certifiable neural network estimation. arXiv preprint arXiv:2603.03082, 2026.
  48. [50]Hamza Tariq and Adeel Akhtar. Dual path following for an aerial manipulator under state and input constraints via transverse feedback linearization. arXiv preprint arXiv:2611.01442, 2026.
  49. [51]Junlin Wu, Andrew Clark, Yiannis Kantaros, and Yevgeniy Vorobeychik. Neural Lyapunov control for discrete-time systems. In Advances in Neural Information Processing Systems, volume 36, pages 2939–2955, 2023.
  50. [52]Bai Xue, Naijun Zhan, and Yangjia Li. A characterization of robust regions of attraction for discrete-time systems based on Bellman equations. IFAC-PapersOnLine, 53(2):6390–6397, 2020.
  51. [53]Bai Xue, Qiuye Wang, Naijun Zhan, Shijie Wang, and Zhikun She. Synthesizing robust domains of attraction for state-constrained perturbed polynomial systems. SIAM Journal on Control and Optimization, 59(2):1083–1108, 2021.
  52. [54]Lujie Yang, Hongkai Dai, Zhouxing Shi, Cho-Jui Hsieh, Russ Tedrake, and Huan Zhang. Lyapunov-stable neural control for state and output feedback: A novel formulation. In International Conference on Machine Learning (ICML), volume 235 of Proceedings of Machine Learning Research, pages 56033–56046, 2024.
  53. [55]Dongjie Yu, Haitong Ma, Shengbo Eben Li, and Jianyu Chen. Reachability constrained reinforcement learning. In International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 25636–25655, 2022.arxiv.org/abs/2205.07536
  54. [56]Pietro Zanotta, Panos Stinis, and Ján Drgoňa. SCORE: Statistical certification of regions of attraction via extreme value theory. arXiv preprint arXiv:2603.29658, 2026.

Paper details

vlnc:W_tf43q6bt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents