- The paper introduces advanced modeling frameworks combining physics-based and data-driven approaches to accurately capture nonlinear behaviors like viscoelasticity and hysteresis in DEAs.
- It describes robust control strategies from open-loop feedforward to closed-loop feedback and compound methods that improve precision in soft actuation applications.
- The review highlights self-sensing techniques using both model-based estimation and neural networks, while emphasizing challenges in high-bandwidth sensorless control.
Modeling, Control, and Self-sensing of Dielectric Elastomer Soft Actuators
Introduction
Dielectric elastomer actuators (DEAs) are soft transducers exhibiting high energy density, large strain, low weight, and fast response, positioning them as central components in applications such as soft robotics, haptic devices, and vibration isolation. However, DEAs present intrinsic electromechanical complexities—strong viscoelasticity, nonlinear rate-dependent hysteresis, and vibrational dynamics—making their modeling, high-fidelity control, and integrated self-sensing nontrivial. The reviewed article provides a comprehensive synthesis of the literature on DEA modeling, control methodologies, and self-sensing strategies, along with a critical discussion of open challenges and future directions (2604.17199).
Electromechanical Behaviors of DEAs
The mechanical response of DEAs is determined by a set of tightly coupled nonlinearities:
- Viscoelastic Creep: DEA materials exhibit classic polymeric time-dependent drift under constant excitation. Creep can lead to displacement errors that exceed 90% of range for commercial VHB elastomers at high frequency in the absence of compensation.
- Hysteresis: The input-output relation is characterized by amplitude- and rate-dependent multi-valued maps with pronounced memory effects, manifesting as non-smooth nonlinearities.
- Vibrational Dynamics: Combination of low mechanical and electrical damping can result in underdamped resonances. The resonance landscape is highly sensitive to both the elastomer/electrode composition and geometric configuration.
These effects complicate both open-loop prediction and feedback control, requiring sophisticated modeling and compensation methodologies.
Modeling Approaches
Physics-based Modeling
The state-of-the-art physics-based DEA modeling involves three explicit submodels: an electrical model (accounting for capacitance and resistances—including strain-dependence), a mechanical constitutive model (addressing hyperelasticity and viscoelasticity), and an electromechanical coupling model (often via Maxwell stress frameworks and continuum mechanics energy formulations). Hyperelastic models based on Yeoh, Ogden, and Gent formulations are foundational, but dynamic modeling additionally mandates multi-element (generalized) Maxwell and Kelvin-Voigt approaches for quantitative description of rate dependence and large-strain behavior.
Recent results demonstrate that combining a visco-hyperelastic Kelvin–Maxwell model with frequency-dependent parameters yields robust prediction of sweep-frequency and amplitude-varying responses. Physics-based approaches are fundamentally constrained by system identification/practical tractability—typically yielding high accuracy only in narrow operation domains.
Phenomenological Modeling
Phenomenological models, including polynomial-based block structures, Hammerstein cascades, and neural network surrogates (e.g., LSTM, NARX, GRU), are widely adopted for controller synthesis due to their flexibility. Hysteresis is typically represented via operator-based structures (Preisach, Prandtl-Ishlinskii (P-I)), differential equation-based models (Duhem, Bouc-Wen), and recently, by data-driven architectures capable of capturing high-order nonlinearity and memory effects. Modified P-I models and advanced neural networks now yield sub-mm RMS error under frequency sweeps to 10 Hz, but their generalizability across DEA platforms and wide operation envelopes remains limited.
Control Strategies
Open-loop and Feedforward Control
Open-loop control is fundamentally impeded by unmodeled nonlinearities and time variance, with model-based feedforward inversion and compensation only feasible for well-characterized cases. Feedforward techniques address quadratic input and linear dynamics (via model inversion and inertial links), viscoelastic creep (using transfer function compensation), and hysteresis (inverse Preisach or P-I model, analytical or direct surrogate inversion).
Closed-loop and Feedback Architectures
Closed-loop methodologies remain essential to achieve high-precision tracking under modeling uncertainties and disturbances. PID, robust LMI-based control, and intelligent model-free techniques (fuzzy logic, neuro-fuzzy) are predominantly suited for low-frequency regimes. Sliding mode control (SMC), including adaptive and proxy-based variants, are utilized to mitigate the impact of unmodeled nonlinearities and to offer robustness to variable loads.
Feedback architectures have been further extended with reinforcement learning (Deep Q-learning) and adaptive observers (e.g., extended Kalman filtering, neural state estimation) to enable data-driven operation with minimal a priori model specification.
Compound Feedforward-Feedback Control
For high-precision/high-frequency applications, compound control methods synthesize feedforward compensation for identified blocks (hysteresis, creep, dynamic response) with real-time feedback (PID/PI, SMC, fuzzy logic, or learning-based controllers) that reject remaining uncertainties. Model predictive control, iterative learning, and MRAC-based adaptive schemes have demonstrated improved bandwidth and disturbance rejection but remain nontrivial to scale to large stroke or highly rate-dependent operation.
Self-Sensing Techniques
Exploiting the natural dependence of DEA capacitance on strain, self-sensing methodologies enable sensorless state estimation—crucial for compact integration in soft robotics. Self-sensing is realized through two primary routes:
- Model-Based Capacitance Estimation: Electrical circuit modeling (typically RC or RC-parallel with strain-dependent components) combined with high-frequency probing and signal identification (impedance analysis, recursive least squares, EKF) allows displacement reconstruction from electrical measurements. Capacitance-displacement maps are established analytically or through polynomial/physical fitting. This approach is sensitive to noise and increases feedback delay.
- Data-Driven Techniques: Neural surrogates and regression (BP neural networks, NARX, polynomial fitting) are trained to directly map electrical observables to displacement, bypassing explicit electrical parameter estimation. While highly flexible, these are heavily reliant on exhaustive data and lack generalizable physical interpretability.
Integrated self-sensing has been successfully deployed in closed-loop DEA control, with displacement estimation errors as low as <0.7 mm in slow regime, but increasing significantly with frequency and controller gain due to delay and noise amplification—leading to non-negligible degradation compared to direct sensor feedback.
Implications and Future Directions
The tractability of high-bandwidth, high-precision DEA modeling, control, and self-sensing remains a limiting factor for wide-scale adoption in advanced soft robotic manipulators, biomedical devices, and adaptive haptics. Three main research avenues are identified:
- Generalized Rate-Dependent Hysteresis Models: Accurate, invertible, and computationally tractable hysteresis models covering the entire DEA operation envelope remain elusive. Unified frameworks that hybridize physics-based insights with data-driven adaptivity are needed.
- High-Bandwidth Sensorless Control: Achieving closed-loop performance matching sensor-based architectures at high frequency and amplitude, including in multi-DOF and under-actuated configurations, is an open problem requiring integration of advanced real-time identification, adaptive control, and robust estimation under adversarial noise.
- Scalability to Multi-DOF Systems: Current methods are largely validated on single-DOF systems; extension to multi-DOF DEAs must account for cross-axis coupling, compounded nonlinearity, and increased computational complexity.
Adaptive and learning-based methods, including reinforcement learning and neural surrogate modeling, are expected to become increasingly prominent, but require rigorous guarantees for safety, stability, and interpretability.
Conclusion
The paper provides a substantive technical account of the state of DEA modeling, control, and self-sensing, rigorously cataloging both modeling architectures and control methodologies. While significant progress has been made, the complexity of DEA nonlinear behaviors continues to present fundamental theoretical and practical challenges—particularly for high-performance, multi-DOF, and sensorless applications. Future work is expected to merge model-based insight, real-time adaptive schemes, and data-driven advances to achieve robust, scalable solutions, accelerating the deployment of DEA-based soft robotic systems (2604.17199).
