Teleoperated finger gaiting skill via gaussian process residual learning

A method for retargeting a motion of a hand to a multi-fingered robot hand comprising:

• determining a residual between an optimization-based retargeting method, which minimizing an error between sensed vectors between points on the hand and scaled vectors between same points on the multi-fingered robot hand, and a labeled configuration of the multi-fingered robot hand using a predefined set of hand-labeled calibration poses, wherein determining a residual comprises using a Gaussian Process (GP) residual learning method to determine the residual, wherein using a Gaussian Process (GP) residual learning method comprises:
• collecting a number of paired configurations of the multi-fingered robot hand and the hand using the predefined set of hand-labeled calibration poses, D={(qhi, qri)}i∈[c];
• determining a shared set of fingers between the multi-fingered robot hand and the hand, F;
• determining a subset of hand joints, q[f] associated with each of the shared set of fingers f∈F;
• and
• determining hyperparameters of a multi-output Gaussian Process (GP) to regress the residual between qo*(qhi)[f] and qi [f] for each f∈F using D;
• and
• using an optimization-based assistive function for active fingers of the multi-fingered robot hand to maintain contact against an object during teleoperation with the multi-fingered robot hand, wherein the optimization-based assistive function uses distance constraints among contact points of the multi-fingered robot hand to allow free finger movements in a null space while providing secure contacts against the object in the multi-fingered robot hand.