There are at least two Asimov stories that I recall when a robot “harms” a human.
In the bicentennial man at the very beginning the surgeon robot refuses to perform the procedure as it would harm Andrew, however he then proceeds once Andrew tells him he’s not a human being.
In the second, Evidence, Stephen Byerley is implied to have been attacked by a robot programmed to do so, with Susan Calvin implying that’s only possible if the robot is also attacking a robot.
In both cases it is implied that a robot can attack another robot but that it needs to know it is and that for humanoid robots they may not and need to be told.
This opens the loophole possibility of having a robot harm a human thinking it is a humanoid robot.
Has that been addressed by Asimov either through a story or directly in writing anywhere?