In the human-robot collaborative interaction mode, when you throw the ball from your hand, the binocular vision system locks onto the ball with a response speed of five milliseconds. The IMU-based attitude estimation algorithm calculates the return trajectory in real time, ensuring that the throwing force remains strictly within a safe range. When the mutual throwing mode is activated, the system initiates a dynamic priority strategy: as the right arm catches a new ball, the left arm passes the handball laterally to the right workspace. This movement effectively trains the user's hand-eye coordination. In the three-ball collaborative test, the vision system simultaneously tracks multiple moving targets. The motion planner divides the workspace into three safety zones, compressing the decision cycle to 0.3 seconds and achieving a throw-and-catch frequency of 3 Hz. This multi-tasking mode can significantly enhance cognitive flexibility in stroke patients, effectively supporting their upper-limb rehabilitation training.