Geometry-aware Robot Manipulability Transfer

Humans naturally vary their body posture in order to quickly move or apply forces along specific directions. Such posture changes are strongly linked to the specific requirements of the task at hand, and therefore play a relevant role on task performance. Posture variation also has a significant role in robot manipulation (e.g., pushing/pulling objects, reaching tasks), where manipulability arises as a useful criterion to analyze and control the robot dexterity as a function of its joint configuration. In this context, this paper introduces a novel framework for transferring manipulability ellipsoids to robots. This framework is first built on a probabilistic learning model that allows for the geometry of the symmetric positive definite manifold to encode and retrieve appropriate manipulability ellipsoids. This geometry-aware approach is later exploited for designing a manipulability tracking controller inspired by the classical inverse kinematics problem in robotics. Experiment in simulation with planar robot arms validate the feasibility of our manipulability transfer framework.