NAVER LABS

Tendon-driven actuation allows for light and compact manipulator designs with enhanced safety features. One of the key challenges in model-based control of tendon-driven robots lies in increasing complexity of the dynamic model due in large part to the hard-to-model components like nonlinear dynamic deformation of the tendon and complex friction behavior. While some of the data-driven modeling approaches, e.g., neural networks, have proven their flexibility and strong function approximation capabilities, and perhaps more importantly free one from dealing with complex and often error-prone mechanics models, they usually do not generalize well to diverse tasks, and also do not offer the needed intuitive understanding or predictive power of traditional mechanics-based models. In this paper, we present a hybrid modeling approach for complex tendon-driven robots, which effectively complement the limitations of pure physics-based and data-driven learning-based approaches. Rigid multibody equations of motion are augmented with (i) a configuration-dependent viscous-Coulomb friction model and (ii) a recurrent neural network that captures the tendon dynamics, and estimates link joint angles from the motor positions, velocities, and torques. Experiments involving a two-dof tendon-driven parallel wrist mechanism and the 7-dof AMBIDEX tendon-driven manipulator, validate the performance advantages of our hybrid model-based control framework.

>https://www.sciencedirect.com/science/article/abs/pii/S0957415820300787