A ProMPs-Based Framework for Robot Variable Impedance Skill Acquisition

Robustness and adaptability are key requirements for robots to interact with the physical environment, and impedance/admittance control play an important role in solving this problem. The robot learning of the variable impedance strategy based on dual Dynamic movement primitives (DMPs) has achieved good results; however, the single demonstration limitation inherent in DMPs makes it necessary to handcraft routine multiple demonstrations, inevitably incorporating personal preferences. We propose a framework based on probabilistic movement primitives (ProMPs) to address this problem in a data-driven manner. Its strategy representation in probabilistic form captures the whole information from human experts’ demonstrations in an objective manner rather than a subjective one. In addition, we have added critical moment points and endpoint moment-over-point constraints to the contours of the strategy expressions for position and stiffness, thus alleviating the problem of large endpoint deviations that has been a problem with conventional ProMPs. In this way, both flexibility of learning and precision of implementation in variable impedance skill transfer can be guaranteed. Finally, simulation experiments demonstrate the effectiveness and scalability of our proposed method.