Improving human-robot physical interaction with inverse kinematics learning

Abstract

More and more collaborative robots are making it into factory floors. These safe robots are meant to physically interact with human operators in tasks involving handing over objects or behaving as a third hand. Inverse kinematics is a key functionality for this as the robot has to find joint configurations to reach specific task space targets. Standard inverse kinematics libraries can be difficult to manipulate when controlling redundant actuators as the obtained configurations can be suboptimal in terms of naturalness and operator comfort. We describe a learning approach that allows the operator to easily adjust the postures of the robot by online demonstration. The learned inverse kinematics functions is used in conjunction with standard inverse kinematics libraries to improve the generated postures. A user study shows that human-robot face-to-face interaction is improved by the learned inverse kinematics.