Physical Interaction as Communication: Learning Robot Objectives Online from Human Corrections

When a robot performs a task next to a human, physical interaction is\ninevitable: the human might push, pull, twist, or guide the robot. The\nstate-of-the-art treats these interactions as disturbances that the robot\nshould reject or avoid. At best, these robots respond safely while the human\ninteracts; but after the human lets go, these robots simply return to their\noriginal behavior. We recognize that physical human-robot interaction (pHRI) is\noften intentional -- the human intervenes on purpose because the robot is not\ndoing the task correctly. In this paper, we argue that when pHRI is intentional\nit is also informative: the robot can leverage interactions to learn how it\nshould complete the rest of its current task even after the person lets go. We\nformalize pHRI as a dynamical system, where the human has in mind an objective\nfunction they want the robot to optimize, but the robot does not get direct\naccess to the parameters of this objective -- they are internal to the human.\nWithin our proposed framework human interactions become observations about the\ntrue objective. We introduce approximations to learn from and respond to pHRI\nin real-time. We recognize that not all human corrections are perfect: often\nusers interact with the robot noisily, and so we improve the efficiency of\nrobot learning from pHRI by reducing unintended learning. Finally, we conduct\nsimulations and user studies on a robotic manipulator to compare our proposed\napproach to the state-of-the-art. Our results indicate that learning from pHRI\nleads to better task performance and improved human satisfaction.\n