Visual IRL for Human-Like Robotic Manipulation