Design and Control of a Multi-Fingered Soft-Rigid Hybrid Robotic Hand

Abstract

In robot hands, compliance improves the quality of grasps and allows for robustness in contact with the environment, which is why soft robot hands, which are inherently compliant, generate such interest despite being complex to control and model. In prior work, our lab developed a soft-rigid hybrid architecture for a robot finger, with the intention of making a compliant finger that is as easy to control as a rigid robot. This thesis details the work done to take this architecture and develop it into a five-fingered dexterous gripper capable of highly compliant grasping — over several iterations, we create an integrated tendon-driven hand that is robust, maintainable, and inexpensive. We develop a precise controller for the soft-rigid hybrid finger, and extend it for both position and task space control of the hand — additionally we implement variable stiffness control within the controller without the need for additional hardware, via adjusting gain values in the control loop. We test the ability of the hand to complete the full set of human grasping postures, and demonstrate that the soft-rigid architecture enables a high degree of generalization, able to complete 28 of the 33 identified human grasp postures. Additionally, tests illustrate the hand’s advantages in completing traditionally difficult manipulation tasks such as picking up thin deformable objects (such as a dollar bill or folding cloth) as well as in interfacing with soft or delicate target objects. We adapt a teleoperation system to map the movements of the robot gripper to a glove worn by a human operator, and evaluate the usability of the hand as a teleoperation target for completing several tasks — we illustrate promising results that the compliance of the hand compensates for operator error and allows for fast completion of tasks requiring environmental or object contact, traditionally difficult tasks for existing rigid robots. Finally, we discuss the use of the teleoperation system to record demonstrations which we then use to train an imitation learning model, utilizing an implementation of denoising diffusion probabilistic models, to complete grasping tasks. We show that our soft-rigid fingers allow a dexterous hand to be trained to perform autonomous grasping with a relatively small set of expert demonstrations, and that the compliance of the physical structure allows for variance in the environment and object position to be compensated for by the physical properties of the hand.