Variable impedance learning control with faster re-learning and reduced initial errors in re-perturbation for robots operating in divergent force fields

Abstract

In recent years, robots with higher levels of sophistication in controlling the mechanical impedance of interaction have received increasing attention. These robots can precisely move and manipulate objects in predetermined situations by exploiting impedance control. However, this requires that the desired impedance of interaction be factored into the control design process beforehand, and interaction impedance is typically not dynamically modulated traditionally. In contrast, humans are skilled at modulating mechanical impedance, mainly mechanical stiffness, in their interactions with their surroundings. This ability is vital for successful interaction with the environment. This human behaviour has inspired various variable impedance learning controllers for tool manipulation. However, humans not only demonstrate the ability to stabilise interaction with a divergent force-field but also demonstrate a few higher-level learning features. Savings is one such feature, and it refers to people being able to learn faster when the same divergent-field is applied again. Retention is another essential feature, allowing humans to initially demonstrate lower error when the same exposure to a divergent field (perturbation) is applied again. These features have practical utility in the robotic application and have not been explored, this paper presents a novel approach to variable impedance learning control for robots inspired by human motor learning mechanisms. We combine iterative learning control, feedforward control, and ideas from neuroscience to synthesise the controller. The proposed method aims to vary impedance to stabilise the interaction with divergent-field while incorporating savings and retention features. Simulation results with a two-link serial chain manipulator demonstrate the proposed method’s efficacy, savings, and retention.