Adaptive Sliding Mode Position-Feedback Controller

• Sliding mode adaptive position-feedback tracking controllers are advanced robust control architectures that combine sliding mode invariance with adaptive compensators for precise trajectory tracking in uncertain nonlinear systems.
• They employ a composite control law integrating equivalent control, switching terms, and adaptive/NN components to mitigate uncertainties, chattering, and disturbances.
• This approach has proven effective in diverse applications such as industrial robotics, underwater vehicles, and soft manipulators with rapid error convergence and strong disturbance rejection.

A Sliding Mode Adaptive Position-Feedback Tracking Controller is a robust nonlinear control architecture integrating classical sliding mode control (SMC) with adaptive mechanisms—often neural or fuzzy network compensators—to achieve high-precision trajectory tracking in systems subject to significant model uncertainties and unknown external disturbances. This paradigm extends core SMC principles—robustness against matched uncertainties, fast error convergence, and strong invariance properties—by coupling them with online adaptation and learning to handle time-varying or unmodeled effects without sacrificing closed-loop stability. Across its many realizations, this design has demonstrated superior rejection of bounded disturbances, chattering attenuation, and high tracking accuracy in complex, nonlinear multi-degree-of-freedom systems, particularly in industrial robotic manipulators, soft continuum arms, underwater vehicles, and flexible large-scale structures (Le et al., 8 Jan 2025, Mustafa et al., 2024, Kazemipour et al., 2021, Li et al., 2022).

1. Feedback Structure, System Modeling, and Uncertainty Framework

Sliding mode adaptive position-feedback tracking controllers are grounded in a measured output (typically joint, link, or global position), often with full-state feedback or robust state estimation. The generic plant can be abstracted as a nonlinear affine model with parameter uncertainties and additive bounded external disturbances:

$$M(q)\,\ddot{q} + C(q,\dot{q})\,\dot{q} + F(q)\,\dot{q} + G(q) = T - f_{\text{ext}}$$

where $q$ represents configuration, $M$ the inertia matrix, $C$ the Coriolis/centrifugal matrix, $F$ friction, $G$ gravity, $T$ the control input, and $f_{\text{ext}}$ an unknown but bounded disturbance (Le et al., 8 Jan 2025). States are $x = [q; \dot{q}]$, the measured output is $y = q$, and the control objective is to drive the tracking error $e(t) = q_d(t) - q(t)$ (with reference $q_d$) to zero.

Model parameter uncertainties (masses, inertias, etc.) are assumed bounded and/or slowly varying; $f_{\text{ext}}$ is only known to satisfy $\|f_{\text{ext}}(t)\| \leq \Delta$.

2. Sliding Surface Design

The design adopts a sliding surface of first or higher order synthesizing position and (possibly) velocity errors:

$$s(t) = \dot{e}(t) + \Lambda e(t), \quad \Lambda = \operatorname{diag}(\lambda_1, ..., \lambda_n) > 0$$

or, for more aggressive convergence and finite-time behavior, a terminal or fractional-order surface,

$$s(t) = \dot{e}(t) + \Lambda |e(t)|^{\alpha} \text{sign}(e(t)), \quad 0.5 < \alpha < 1$$

as in soft/continuum manipulation and fixed-time satellite tracking (Kazemipour et al., 2021, Sahoo et al., 13 Feb 2025). These surfaces ensure that when $s = 0$, the tracking error dynamics reduce to an exponentially or finitely convergent ODE.

In high-order or discrete implementations (e.g., sampled automotive or flexible mechanical systems), composite sliding variables are constructed to ensure convergence of both the tracking error and its derivatives, e.g.,

$$\xi(k) = s(k+1) + \beta s(k), \quad s(k) = x(k) - x_d(k)$$

with $\beta > 0$ (Amini et al., 2018, Amini et al., 2017).

3. Controller Synthesis: Equivalent, Switching, and Adaptive/NN Terms

The control law is structured as a composition of three main elements:

1. Equivalent Control ($u_{\text{eq}}$): A model-based compensation that linearizes the nominal nonlinear plant along the sliding manifold—generally,

$$u_{\text{eq}}(t) = M(q) \ddot{q}_d + C(q,\dot{q}) \dot{q}_d + G(q)$$

or an explicit regressor-based estimate,

$u_{\text{eq}} = Y(q,\dot{q},\cdots)\hat{\theta}$

2. Switching/Robustifying Term ($u_{\text{sw}}$): A discontinuous injection of sufficient amplitude to dominate bounded uncertainties,

$u_{\text{sw}}(t) = -K\,\text{sign}(s)$

where $K$ is selected such that $K > \text{bound}(\Delta_{\text{residual}})$. For chattering reduction, the discontinuity is mollified via a boundary layer: $$\text{sign}(s) \rightarrow \text{sat}(s/\epsilon), \hspace{1em} \epsilon > 0$$

For higher-order or continuous-time/digital systems, second-order (super-twisting, terminal, integral) switching laws or adaptive continuous correction terms are introduced to overcome limitations of data sampling, model discrepancies, and chattering (Wang et al., 2020, Li et al., 2022).

3. Adaptive/Neural Compensation ($u_{\text{NN}}$): To capture the residual nonlinear uncertainty and adaptively augment robustness, a feedforward neural network or fuzzy compensator is frequently deployed:

$$u_{\text{NN}}(t) \approx \hat{W}^T \phi(q, \dot{q}, s)$$

with real-time weight adaptation via gradient-driven laws: $\dot{\hat{W}} = \Gamma\,\phi(q, \dot{q}, s) s^T$ or, for fuzzy systems, TSK rule-based interpolation and adaptation: $\dot{\hat{\mathbf{D}}} = -\gamma\,s\,\Psi(\hat{u})$

This architecture yields a final law: $$u(t) = u_{\text{eq}}(t) + u_{\text{sw}}(t) + u_{\text{NN}}(t)$$ ensuring that the controller maintains performance despite substantial uncertainty and environmental disturbance (Le et al., 8 Jan 2025, Zuo et al., 9 Sep 2025, Bessa, 2022).

4. Adaptation and Learning Mechanisms

Adaptation is realized through online update of parameter estimates (be they plant or NN weights), typically governed by Lyapunov-stable laws driven by the sliding variable(s):

• Parameter adaptation:

$\dot{\hat{\theta}} = \Gamma\,Y^T\,s$

with $\Gamma > 0$ an adaptation rate, and $Y$ the regression/basis matrix.

• NN weight update:

$\dot{\hat{W}} = \Gamma\,\phi s^T$

• Disturbance/uncertainty bound adaptation:

$\dot{\hat b} = \Psi\,|s|$

• Reinforcement learning–based gain adaptation: as in fixed-time SMC for satellite docking, where a neural actor–critic module tunes the switching slopes to optimize the value function in the presence of unknown system dynamics (Sahoo et al., 13 Feb 2025).

These mechanisms allow the controller to dynamically compensate both slow and fast-varying uncertainties, time-varying disturbances, and unmodeled actuator or payload effects, preserving boundedness and guaranteeing asymptotic (or fixed-time) error convergence.

5. Stability and Lyapunov Analysis

Robust stability and convergence are formally established via composite Lyapunov functions combining energy on the sliding manifold and parameter/weight errors:

$$V(s, \tilde{W}) = \frac{1}{2} s^T M(q) s + \frac{1}{2} \text{Tr}\{(\tilde{W})^T \Gamma^{-1} \tilde{W}\}$$

or more generally,

$$V = \frac{1}{2}s^T K s + \frac{1}{2}\tilde\theta^T \Gamma^{-1} \tilde\theta + \frac{1}{2}\tilde b^T \Psi^{-1} \tilde b$$

where $\tilde{W} = \hat{W} - W^*$, $\tilde\theta = \hat\theta - \theta^*$, and $\tilde b = \hat b - b^*$. Under the dynamics induced by the composite controller and adaptation law, the time derivative satisfies

$$\dot{V} \leq -s^T K_{\text{sw}}\,\text{sat}(s/\epsilon) + |s|^T \Delta_{\text{residual}} \leq -\eta\|s\| + c$$

for some $\eta > 0$, $c$ small. When controller gains are chosen to dominate residual uncertainties, $s \to 0$ in finite time and $e,\dot{e} \to 0$ asymptotically (Le et al., 8 Jan 2025, Kazemipour et al., 2021, Mustafa et al., 2024, Amini et al., 2018, Li et al., 2022).

In cases employing higher-order or fixed-time laws, the Lyapunov analysis is extended to demonstrate uniform ultimate boundedness or convergence to the manifold in a time independent of initial conditions (Wang et al., 2020, Sahoo et al., 13 Feb 2025).

6. Practical Implementation and Performance

The architecture is modular and hardware-agnostic, applicable from industrial 3-DOF cylindrical and 6-DOF serial robots, flexible marine cranes, and AUVs, to soft continuum manipulators. Standard block-diagram flow:

[q_d, q̇_d, q̈_d] → e, ē → sliding s → [u_eq, u_sw, u_NN] → Plant → q, q̇ → s, NN adaptation

Measured performance metrics, validated via simulation and (in some works) physical experiments, include:

• Asymptotic/finite-time convergence: e(t), ē(t) → 0; typical settling times <0.5 s (cylindrical manipulator), 1.3 cm steady-state error (soft arm), <0.02 rad error (AUV formation) (Le et al., 8 Jan 2025, Kazemipour et al., 2021, Li et al., 2022).
• Robustness: performance maintained under step/unbounded parameter changes and external disturbances (force/torque injection).
• Chattering reduction: boundary layers and NN/fuzzy compensation shrink error bumps/reachability artifacts to sub-millisecond or sub-centimeter regimes.
• Disturbance rejection: rapid recovery (error suppression in ~0.1 s) upon perturbation.
• Adaptive efficiency: parameter/weight estimates converge in real time; adaptation rates and gain selection trade off between convergence speed and noise/chattering (Le et al., 8 Jan 2025, Zuo et al., 9 Sep 2025).
• Comparative advantage: supersedes conventional SMC, fuzzy-SMC, PID, and classical LQR in tracking accuracy, robustness, and response speed (e.g., MSE improvement orders of magnitude; chattering almost eliminated) (Zuo et al., 9 Sep 2025).

7. Application-Specific Variants and Extensions

Numerous domain-specific variants and applications have been explored:

Domain / Application	Key Features / Adaptations	Reference
Cylindrical manipulators	NN-adaptive SMC, high-precision 3D printing	(Le et al., 8 Jan 2025)
Discrete-time platforms	Second-order DSMC, sampling & quantization robustness	(Amini et al., 2018, Amini et al., 2017)
Underactuated/flexible cranes	Hierarchical sliding, global stability, NN comp.	(Zuo et al., 9 Sep 2025)
Soft continuum robots	Terminal sliding surface, dual adaptation	(Kazemipour et al., 2021)
Marine/AUV formation	Super-twisting, adaptive switching, bounded flow dist.	(Li et al., 2022)
Fractional/fixed-time SMC	Fractional-order/adaptive super-twisting, RL/NN slope tuning	(Chen et al., 2021, Sahoo et al., 13 Feb 2025)
Nonlinear dead-zone compensation	Fuzzy-SMC, adaptive laws, zero-order TSK	(Bessa, 2022)

Each variant selects the sliding surface, adaptive architecture, and chattering-mitigation strategy to match domain constraints (e.g., sampling intervals, actuator nonlinearities, noise). Integration with state observers, super-twisting algorithms, or fuzzy/NN learning is common when direct measurement or model information is incomplete.

---

A Sliding Mode Adaptive Position-Feedback Tracking Controller offers a flexible, theoretically grounded approach to robust tracking in complex nonlinear systems, merging the invariance properties of sliding mode with modern online adaptation schemes to secure high tracking accuracy, fast response, and strong disturbance rejection, as systematically validated in robotic, marine, automotive, and aerospace systems (Le et al., 8 Jan 2025, Zuo et al., 9 Sep 2025, Amini et al., 2018, Li et al., 2022).