Model-guided design of dielectric elastomer actuators(DEAs)is essential for enabling their application in soft robotics.However,current modeling methods primarily rely on the finite element method(FEM),which suffers f...Model-guided design of dielectric elastomer actuators(DEAs)is essential for enabling their application in soft robotics.However,current modeling methods primarily rely on the finite element method(FEM),which suffers from low computational efficiency.Additionally,the simulation-to-reality(Sim2Real)gap,mainly arising from variations in material properties and manufacturing processes,poses a significant challenge.In this work,we propose a data-driven modeling framework aimed at accurately and rapidly predicting voltage-induced displacements while minimizing the Sim2Real gap.The framework integrates a multi-layer perceptron(MLP)model,which serves as a computationally efficient surrogate for the FEM model,and a cycle-generative adversarial network(CycleGAN)model,which mitigates the Sim2Real gap by leveraging adversarial learning to process both simulation and experimental data.Dimensional analysis is performed to extend the framework's applicability across different DEA scales.The surrogate model delivers global predictions in just 0.8 s,achieving linear coefficients of determination(R^(2))of 0.99106 for release distance prediction and 0.99375 for actuation distance prediction compared to experimental results.Our model can quickly identify the feasible range of biaxial prestretch ratios required for generating the desired deformation,thereby streamlining the design process.Finally,a soft robotic gripper is designed and fabricated,demonstrating versatile object-grasping capabilities.展开更多
This paper presents a learning-based control framework for fast(<1.5 s)and accurate manipulation of a flexible object,i.e.,whip targeting.The framework consists of a motion planner learned or optimized by an algori...This paper presents a learning-based control framework for fast(<1.5 s)and accurate manipulation of a flexible object,i.e.,whip targeting.The framework consists of a motion planner learned or optimized by an algorithm,Online Impedance Adaptation Control(OIAC),a sim2real mechanism,and a visual feedback component.The experimental results show that a soft actor-critic algorithm outperforms three Deep Reinforcement Learning(DRL),a nonlinear optimization,and a genetic algorithm in learning generalization of motion planning.It can greatly reduce average learning trials(to<20 of others)and maximize average rewards(to>3 times of others).Besides,motion tracking errors are greatly reduced to 13.29 and 22.36 of constant impedance control by the OIAC of the proposed framework.In addition,the trajectory similarity between simulated and physical whips is 89.09.The presented framework provides a new method integrating data-driven and physics-based algorithms for controlling fast and accurate arm manipulation of a flexible object.展开更多
基金supported by the National Key Research and Development Program of China(Grant No.2023YFB4706500)the National Natural Science Foundation of China(Grant Nos.52275026,T2293725)+1 种基金the Shanghai Rising-Star Program(Grant No.24QA2704500)the Natural Science Foundation of Shanghai(Grant No.23ZR1427800)。
文摘Model-guided design of dielectric elastomer actuators(DEAs)is essential for enabling their application in soft robotics.However,current modeling methods primarily rely on the finite element method(FEM),which suffers from low computational efficiency.Additionally,the simulation-to-reality(Sim2Real)gap,mainly arising from variations in material properties and manufacturing processes,poses a significant challenge.In this work,we propose a data-driven modeling framework aimed at accurately and rapidly predicting voltage-induced displacements while minimizing the Sim2Real gap.The framework integrates a multi-layer perceptron(MLP)model,which serves as a computationally efficient surrogate for the FEM model,and a cycle-generative adversarial network(CycleGAN)model,which mitigates the Sim2Real gap by leveraging adversarial learning to process both simulation and experimental data.Dimensional analysis is performed to extend the framework's applicability across different DEA scales.The surrogate model delivers global predictions in just 0.8 s,achieving linear coefficients of determination(R^(2))of 0.99106 for release distance prediction and 0.99375 for actuation distance prediction compared to experimental results.Our model can quickly identify the feasible range of biaxial prestretch ratios required for generating the desired deformation,thereby streamlining the design process.Finally,a soft robotic gripper is designed and fabricated,demonstrating versatile object-grasping capabilities.
基金supported in part by the Brødrene Hartmanns(No.A36775)Thomas B.Thriges(No.7648-2106)+1 种基金Fabrikant Mads Clausens(No.2023-0210)EnergiFyn funds.
文摘This paper presents a learning-based control framework for fast(<1.5 s)and accurate manipulation of a flexible object,i.e.,whip targeting.The framework consists of a motion planner learned or optimized by an algorithm,Online Impedance Adaptation Control(OIAC),a sim2real mechanism,and a visual feedback component.The experimental results show that a soft actor-critic algorithm outperforms three Deep Reinforcement Learning(DRL),a nonlinear optimization,and a genetic algorithm in learning generalization of motion planning.It can greatly reduce average learning trials(to<20 of others)and maximize average rewards(to>3 times of others).Besides,motion tracking errors are greatly reduced to 13.29 and 22.36 of constant impedance control by the OIAC of the proposed framework.In addition,the trajectory similarity between simulated and physical whips is 89.09.The presented framework provides a new method integrating data-driven and physics-based algorithms for controlling fast and accurate arm manipulation of a flexible object.
