Output-Feedback Integral Sliding-Mode Extremum Seeking with Super-Twisting Differentiation

summary

The gist

This paper develops a dither-free output-feedback extremum-seeking controller for pre-stabilized nonlinear dynamic systems with unknown steady-state performance maps The proposed realization succeeds…

In short

The research developed a dither-free output-feedback controller for extremum seeking in nonlinear systems with unknown performance maps. It uses an integral sliding mode mechanism to enforce a constant search velocity, combined with super-twisting differentiation and a ramp-corrected coordinate to achieve practical convergence and explicit bounds on transients. The method successfully outperformed other techniques across 200 test cases.

Key concepts

Dynamic Extremum Seeking Formulation
This is the problem of finding the optimal value of a system parameter ($ heta$) that minimizes an unknown steady-state performance function ($J$). The goal is to steer $ heta$ toward its unknown minimum ($ heta^ ext{*} $) by using measured performance data and applying control inputs.
Integral Sliding Mode Mechanism
This mechanism forces the rate of change of the applied parameter to equal a specific, non-zero search velocity after initial capture. This ensures a consistent search motion even when facing bounded disturbances in the system dynamics, leading to robust parameter tracking.
Super-Twisting Differentiation
This technique is used to estimate the time derivatives of signals (like the performance gradient) in finite time. It provides a smooth, exact derivative estimate within a defined state-parameter tube, which is crucial for accurate control and gradient calculation.
Ramp-Corrected Moving-Equilibrium Coordinate
This coordinate system is designed to cancel out the first-order ramp term caused by the plant dynamics. By using this coordinate, the controller can explicitly derive bounds on transient behaviors like flow, reversal, and recovery time, guaranteeing local practical convergence.

Terminology used across episodes

This episode discusses

The paper

Output-Feedback Integral Sliding-Mode Extremum Seeking with Super-Twisting Differentiation · Read on arXiv

Department of Mechanical Engineering, University of Maryland

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Output-Feedback Integral Sliding-Mode Extremum Seeking with Super-Twisting Differentiation".

Dev: Here's the gist of the paper. This paper develops a dither-free output-feedback extremum-seeking controller for pre-stabilized nonlinear dynamic systems with unknown steady-state performance maps The proposed realization succeeds in all 200 held-out cases and attains the smallest upper-tail objective gap under the stated tuning protocol Dynamic Extremum Seeking Formulation The dynamic extremum-seeking problem involves steering a tunable parameter to a neighborhood of an unknown minimizer of an unknown steady-state input–output characteristic The objective is to steer θ to a neighborhood of the unknown minimizer θ⋆ of J, using the measured performance and the available applied parameter Assumptions define conditions on the plant dynamics and the performance map, including smooth uniformly stable equilibria and a detectable strict minimum Assumption 2 specifies that J has a unique minimizer θ⋆ in int Θ, satisfying sgn J′(θ) = sgn(θ − θ⋆) (θ, θ⋆) Integral Sliding Mode Mechanism The integral sliding mode enforces equality between the applied parameter rate and a prescribed nonzero search velocity after finite-time capture under bounded matched rate-channel perturbations The controller implements θ˙ = uθ + dθ, where uθ is the outer-loop rate command, and the commanded search velocity satisfies v(t) ∈ (−ρ, +ρ), ρ > 0 Proposition 1 shows that if kI > dθ, the reaching time TI from an arbitrary initial sliding error satisfies TI ≤ s(0) kI − dθ Super-Twisting Differentiation and Ratio A super-twisting differentiator provides finite-time estimates of signal derivatives The flow-time gradient surrogate is bg:= w/v, where w is the derivative estimate supplied to the ratio surrogate The denominator is known and nonzero because v = ρ A deadbanded hybrid relay establishes a local practical convergence radius Practical Convergence and Bounds The analysis yields a ramp-corrected moving-equilibrium coordinate that cancels the first-order ramp-induced plant term and yields explicit flow, reversal, recovery, and dynamic-remainder bounds Lemma 1 proves that for any flow interval beginning at ta, Vη(t) ≤ e−ˡh(t−ta)Vη(ta) + βηρ4 / (4), which implies a recovery time τr ≥ 1 / lħ log+ (V+η(CV − βηρ4)) Theorem 1 establishes that every initial condition generates a forward-complete hybrid Filippov trajectory, and the pair (x(t), θ(t)) remains in Z Validation and Performance The proposed controller succeeds in all 200 held-out cases on the 80-state tubular-reactor benchmark The matched-rate sweep associates No-ISM degradation with increasing applied-rate mismatch, followed by nearstall motion and direction reversal The causal ISM–STW succeeds in all 100 cases at every κ, its 90th-percentile gap remains below 3.6 × 10−4 The proposed method has the highest success rate and the smallest 90th-percentile gap among all tested methods An output-feedback ISM–STW extremum seeker was developed for stabilized nonlinear systems with unknown equilibrium maps The resulting practical parameter-error and gradient bounds separate dynamic lag, sensing, deadband, and dwell effects The causal realization succeeds in all 200 held-out cases and attains the smallest upper-tail gap The matched-rate sweep associated No-ISM degradation with increasing rate mismatch, followed by nearstall motion and direction reversal The proposed method succeeded for every tested threshold pair and initialization level References Al-Ghadiri, T.A., Al-Samarraie, S.A., Humaidi, A.J., Al Mhdawi, A.K., 2026 Aminde, N.O., Oliveira, T.R., Hsu, L., 2024 Angulo, M.T., 2015 Ariyur, K.B., Krstic, M., 2003 Dürr, H.B., Stankovic, M.S., Ebenbauer, C., Johansson, K.H., 2013 Efimov, D., Fridman, E., 2025 Filippov, A.F., 1988 Ghaffari, A., Krstic, M., Nešic, D., 2012 Hunnekens, B.G.B., Haring, M.A.M., van de Wouw, N., Nijmeijer, H., 2014 Kim, H., 2026 Krstic, M., Wang, H.H., 2000 Levant, A., 1998 Levant, A., 2003 Luo, R., Rodrigues, V.H.P., Oliveira, T.R., Krstic, M., 2026 Moreno, J.A., Osorio, M., 2012 Salsbury, T.I., House, J.M., Alcala, C.F., 2020 Scheinker, A., 2024 Scheinker, A., Krstic, M., 2014 Solís, C.U., Clempner, J.B., Poznyak, A.S., 2019 Solís, C.U., Clempner, J.B., Poznyak, A.S., 2021 Tan, Y., Nešic, D., Mareels, I., 2006 Torres-Zúñiga, I., López-Caamal, F., Hernández-Escoto, H., Alcaraz-González, V., 2021 Trollberg, O., Jacobsen, E.W., 2018 arXiv:2611.08345v1

eess.SY: 5 Nov 2026 Table 4 Method Success Median eJ Q0.9(eJ) Median eθ Median Aθ ------------------ Causal ISM–STW 200/200 1.49×10−4 3.28×10−4 0.0167 0.0108 Rate-informed STW 174/200 9.88×10−5 1.11×10−2 0.0458 2.66×10−3 P-ESC-T 196/200 9.57×10−4 2.61×10−3 0.0458 2.27×10−3 P-ESC-F 28/200 1.02×10−1 2.65×10−1 3 × 10−3 --- Disturbance ratio κ --- --- Q0.9(eJ) --- Causal ISM–STW 3 × 10−4 (at κ=1.2) No-ISM 0.400 (at κ=0.9) P-ESC-T 2.25×10−3 (at κ>1) Table 6 Condition or quantity Value / status ------ qs 9.08 × 10−5 τr / τmin 0.4861 / 0.5000 Local-consistency left / right side 5.0 × 10−8 / 2.5 × 10−6 C∆ 1.19 εtot / Rρ 0.0169 / 0.0169 Parameter-error / gradient bound 0.0219 / 0.0219 TD / TA 3.8476 / 3.8476 Table 5 Quantity Value / status ------ Rate-authority ratio at κmax = 1.3, kI/(κmaxρ) 1.48 Endpoint-LMI margin 4.34 × 10−2 Maximum recovered derivative error 1.12 × 10−3 Table 3 Method Selected values ------ Causal ISM–STW ρ = 0.0107882, τmin = 0.814674, hg = 0.0019813, (λ1, λ2) = (0.36755, 0.11173), TA = 0.940735, kI = 0.02 P-ESC-T (ω, ap, kesc) = (0.582654, 0.0316391, 2.27289), (rh,rl) = (0.0359860, 0.070139), ϕ = −− Rate-informed STW (λc, αc, εc) = (5×10−7, 6.49×10−4, 2.66×10−3) Table 2 Method Global search box ------ Causal ISM–STW ρ ∈

0.004, 0.015: ; τmin ∈

0.7, 4.5: ; hg ∈

0.002, 0.18: P-ESC-T ω ∈

1e-2, 18: ; ap ∈

3e-3, 7e-2: ; kesc ∈

1e-2, 3: , rh ∈

1.5e−2, 0.7: ; rl ∈

0.03, 1.5: ; ϕ ∈

-π, π: Rate-informed STW λc ∈

2e-3, 6.5e-4: ; αc ∈

5×10−7, 1.5×10−3: ; εc ∈

5×10−4, 4 × 10−2: ; λ1c ∈

8e-2, 6.5e-2: ; λ2c ∈

1.2e-3, 3 × 10−1: Figure 3 Figure 3 reports the sweep Table 7 Method Selected values ------ Causal ISM–STW ρ = 0.0107882, τmin = 0.814674, hg = 0.0019813, (λ1, λ2) = (0.36755, 0.11173), TA = 0.94 P-ESC-T (ω, ap, kesc) = (0.582654, 0.0316391, 2.27289), (rh,rl) = (0.0359860, 0.070139), ϕ = −− Rate-informed STW (λc, αc, εc) = (1.865×10−4, 5.09×10−7, 6.49×10−4) Figure 2 Figure 2 shows the empirical CDFs of eJ in the held-out cases Table 3 (Final) Method Success Median eJ Q0.9(eJ) Median eθ Median Aθ ------------------ Causal ISM–STW 200/200 1.49×10−4 (at κ=1.2) 3.28×10−4 Table 5 (Final) Quantity Value / status ------ Rate-authority ratio at κmax = 1.3, kI/(κmaxρ) 1.48 Table 6 (Final) Condition or quantity Value / status ------ qs 9.08 × 10−5 Table 7 (Final) Method Selected values ------ Causal ISM–STW ρ = 0.0107882, τmin = 0.814674, hg = 0.0019813, (λ1, λ2) = (0.36755, 0.11173), TA = 0.94 P-ESC-T (ω, ap, kesc) = (0.582654, 0.0316391, 2.27289), (rh,rl) = (0.0359860, 0.070139), ϕ = −− Rate-informed STW (λc, αc, εc) = (1.865×10−4, 5.09×10−7, 6.49×10−4) Table 3 (Final) Method Selected values ------ Causal ISM–STW ρ = 0.0107882, τmin = 0.814674, hg = 0.0019813, (λ1, λ2) = (0.36755, 0.11173), TA = 0.94 P-ESC-T (ω, ap, kesc) = (0.582654, 0.0316391, 2.27289), (rh,rl) = (0.0359860, 0.070139), ϕ = −− Rate-informed STW (λc, αc, εc) = (1.865×10−4, 5.09×10−7, 6.49×10−4) Table 2 (Final) Method Global search box ------ Causal ISM–STW ρ ∈

0.004, 0.015: ; τmin ∈

0.7, 4.5: ; hg ∈

0.012, 18: P-ESC-T ω ∈

1e-2, 18: ; ap ∈

3e-3, 7e-2: ; kesc ∈

1e-2, 3: , rh ∈

0.015, 0.7: ; rl ∈

0.03, 1.5: ; ϕ ∈

-π, π: Rate-informed STW λc ∈

8e-2, 6.5e-2: ; αc ∈

5×10−7, 1.5×10−3: ; εc ∈

4×10−4, 4 × 10−2: ; λ1c ∈ [8e-2, 6.

Title and authors: Rosa: So we're looking at this paper called "Output-Feedback Integral Sliding-Mode Extremum Seeking with Super-Twisting Differentiation". It sounds pretty technical, right?

Dev: Yeah, it’s definitely deep into nonlinear control. The title tells you a few things immediately: they're using integral sliding mode and super-twisting differentiation to find an extremum where there isn't an explicit map for the performance.

Rosa: It’s about creating this controller that doesn't need a perfect model of the system; it uses what you measure—the performance, the parameter you apply, and what's happening inside the system—to steer that parameter toward its best possible value.

Dev: Exactly. They’re tackling those unknown steady-state performance maps that make standard extremum seeking methods struggle because they don't know exactly how the system behaves at its best point.

Rosa: It suggests a way to do this dither-free, which is a big deal because dither can introduce noise when you're trying to find the minimum.

Dev: And the authors claim this realization works in all two hundred held-out cases they tested on an eighty-state tubular reactor benchmark, and it achieves the smallest upper-tail objective gap under their tuning protocol.

The paper's summary: Rosa: They summarize by saying this controller uses only the measured performance, the applied parameter, and internal states to do its job. It enforces equality between the rate of change of that applied parameter and a specific nonzero search velocity after it captures the error in finite time under certain noise conditions.

Dev: That integral sliding mode part is key because it ensures there's a consistent search motion even when there are matched rate channel perturbations, where you can only control the parameter rate and not directly its change.

Rosa: And then they combine that with a super-twisting differentiator and some ramp-corrected moving-equilibrium coordinate to recover a gradient sign surrogate without needing any demodulation or averaging.

Dev: That’s the part that sounds really clever; using those tools allows them to get exact differentiation through the local state parameter tube, which is what they need for the super-twisting differentiator to work correctly.

Rosa: They also use a ramp-corrected moving-equilibrium coordinate that cancels out that first order ramp induced plant term, and this gives them explicit bounds on flow, reversal, recovery, and dynamic remainder terms.

Dev: The results show that for any flow interval starting at some time t a, the variable V eta is bounded by an exponential decay term plus a term proportional to the fourth power of rho divided by four.

The paper's improvements: Rosa: One of the main improvements they highlight is how they establish explicit bounds on all those things—the flow, the reversal, and recovery times—using this ramp-corrected moving-equilibrium coordinate. They don't just say it converges; they give you a concrete estimate for how long it takes to recover.

Dev: They prove using lemma one that for any flow interval starting at time t a, the variable V eta is less than or equal to an exponential decay term plus beta eta rho four over four.

Rosa: Which then implies a recovery time τ r is greater than or equal to one over l h log plus (V plus eta CV minus beta eta rho four). That gives you an explicit lower bound on the recovery time.

Dev: And they also establish theorem one which shows that every initial condition generates a forward complete hybrid Filippov trajectory, and that the pair of system state and parameter remains in the set Z.

Rosa: This is important because it means for any starting point you pick, you are guaranteed to stay within a specific set where the system stays bounded, even with all those disturbances.

Dev: They also found that the normalized ratio error separates a dynamic lag contribution from a sensing-rate contribution, and that a deadbanded relay gives an explicit local practical convergence radius.

Conclusion: Rosa: So to wrap up on this paper, they developed an output-feedback integral sliding mode extremum seeker with super-twisting differentiation for systems with unknown performance maps. They achieved success in all two hundred held-out cases on the eighty-state reactor benchmark and got the smallest upper tail gap under their tuning protocol.

Dev: The paper shows how to separate dynamic lag from sensing rate contribution using the normalized ratio error, and they also found that a deadbanded relay provides an explicit local practical convergence radius.

Rosa: They established bounds on plant transients, flow, reversal, and recovery using a ramp-corrected moving-equilibrium coordinate that separates dynamic lag from sensing rate contribution.

Dev: The causal realization succeeds in all two hundred held-out cases and attains the smallest upper tail gap among all tested methods.

Rosa: For someone just listening to the show, this means they've put together a method that is robust enough to handle these unknown steady-state maps without needing prior knowledge of how the system behaves perfectly at its optimum.

Dev: They also showed that with a matched rate sweep, you don't see any loss of gradient sign resolution when the rate mismatch increases, followed by near stall motion and direction reversal.

Rosa: It’s a solid result for anyone working on real-time control of complex nonlinear systems where you can't just rely on perfect mathematical models.

Taro: From my side, I'm just curious about how it handles the world misbehaving part. If the system reverses direction or stalls, does this controller recover gracefully, or does it just stay stuck in that bound?

Dev: That’s a fair question from Taro. The analysis shows explicit flow and reversal bounds yield a uniform recovery estimate for those difficult situations.

Rosa: So they’ve given you a way to predict the recovery time, even when things get messy in the real world, which is exactly what we need when deploying these kinds of controllers outside the lab.

More episodes

← Home