Adaptive whole-body manipulation in human-to-humanoid multi-contact motion retargeting

Abstract

We propose a controller for loco-manipulation motion retargeting from a human to a humanoid robot. Using this controller, the robot can track complex motions and automatically adapt to elements in the environment that have different physical properties from those that were used to provide the human's reference motion. The multi-contact loco-manipulation problem is formulated as a multi-robot quadratic program (MRQP), which optimizes over the combined dynamics of the robot and any manipulated element in the environment. Our approach maintains a dynamic partition of the robot's tracking links into fixed support contact links, manipulation contact links, and contact-free tracking links. The three sets are repartitioned and re-instantiated as objectives or constraints in the MRQP when contact events occur in the human motion. We present various experiments (bag retrieval, door opening, box lifting) using human motion data from an Xsens inertial motion capture system. We show in dynamics simulation that the robot is able to perform difficult single-stance motions as well as multi-contact-stance motions (including hand supports), while adapting to environment elements of varying inertial properties.