Parameter-Learning Active Disturbance Rejection Controller for a Novel Variable-Stiffness Gripper

Abstract The growing demand for flexible robotic grasping in industry calls for adaptable solutions capable of handling diverse objects across stiffness regimes. We present a novel variable-stiffness gripper with a parallel-guided beam and sliding-block mechanism, enabling continuous stiffness modulation (0.143–2.021 N/mm) without component replacement. However, transmission nonlinearities, friction, stiffness-dependent effects, and, in particular, frequency drift arising from stiffness variations significantly hinder precise force control. To overcome these challenges, we propose a parameter-learning active disturbance rejection control (PL-ADRC) framework, integrating online adaptive parameter identification with a model-based extended state observer for real-time estimation and rejection of disturbances arising from unmodeled dynamics and parametric uncertainties. Experimental results demonstrate the superior performance of PL-ADRC: PL-ADRC reduces the band settling time by 0.13 s compared to model-free active disturbance rejection control (MF-ADRC), limits the steady-state force error to 0.01 N, and exhibits robust adaptability in stiffness modes. It outperforms model-based and model-free methods in fragile-object manipulation (e.g., egg grasping without fracture) and high-noise scenarios, achieving faster stabilization and reduced overshoot. This framework bridges precision and adaptability, advancing safe human–robot collaboration in dynamic industrial tasks.