There are several algorithms for post-processing of kinesthetically demonstrated motions for manipulators learning a task. However, I have failed to find papers/research work on modeling the interaction of the robot with human and accommodating as well as generating the back-driving/active compliance motion.
While the human interference can be sensed using joint torque sensors or change in joint angular position, what are the ways to effectively channelize this sensing to actively move the manipulator such that the human doesn't need to apply any considerable external force.
PS - I understand that the manipulator runs a zero gravity or a zero gravity zero friction controller to keep the manipulator floating while in torque control.
