Perpustakaan Digital ITB

Soft robots constitute a class of robotic systems fabricated from elastic materials that offer inherent compliance and safe human interaction, making them particularly suitable for medical applications. However, their continuous deformable structure gives rise to highly nonlinear dynamics, which presents significant challenges for modeling and control. Existing approaches, including physics-based methods, neural networks, and statistical models, suffer from limiting trade-offs between interpretability, data requirements, and computational cost. This study investigates the application of the Sparse Identification of Nonlinear Dynamics (SINDy) algorithm as a data-driven method to identify interpretable, parsimonious governing equations for a single degree-of-freedom (1-DOF) pneumatically actuated soft robot. Experimental data were acquired from a silicone-based cylindrical soft robot subjected to a linear frequency-sweep chirp excitation signal ranging from 0.04 Hz to 0.087 Hz, bounded within a safe pressure limit of 132 kPa. Kinematic state data were captured using an optical passive marker motion capture system at 500 Hz and synchronized with the measured pressure input. The middle segment angle (??) and its time derivative were selected as the independent state variables, with the remaining segment angles reconstructed via static polynomial mapping. Four SINDy configurations were evaluated by combining two sparse optimizers, Sequential Thresholded Least Squares (STLSQ) and Sparse Relaxed Regularized Regression (SR3), with two feature libraries, a pure polynomial library and a combined polynomial-Fourier library. Validation results demonstrate that all four configurations successfully identified sparse governing equations with R² values exceeding 0.83 for the primary identified state. The SR3 optimizer with a polynomial-Fourier library achieved the highest primary state accuracy with R² = 0.9867, while the SR3 with a pure polynomial library produced the most control-suitable model, retaining only four active terms while maintaining an R² of 0.9825 and an RMSE difference of merely 0.34° relative to the more complex configurations. These results confirm that SINDy is a viable and interpretable modeling framework for soft robot dynamics identification.