Mechanical arm control method and device based on hybrid reinforcement learning and underwater robot

The invention relates to a mechanical arm control method and device based on hybrid reinforcement learning and an underwater robot, and belongs to the technical field of mechanical arm control. A positioning error obtained by making a difference between the expected operation pose and the current operation pose of the mechanical arm serves as the input of a controller based on hybrid reinforcement learning, a control instruction serves as the output of the controller, and closed-loop control is conducted on the mechanical arm through the controller; the mixed reinforcement learning comprises construction of a reward function calculation model, a reward function comprises a position reward item and further comprises a posture reward item, and when the distance between the tail end of the current mechanical arm after action and the tail end of the expected mechanical arm is smaller than or equal to a position error set value, the posture reward item is rewarded; the closer the unit attitude quantity at the tail