A system of collaborative control for uncalibrated robotic arms, grounded in 3D vision guidance, is put forward herein. This addresses three central issues in the control system of industrial robotic arms: inadequacy in human-machine collaboration precision, latency in dynamic response, and deficiency of safety constraints. Lightweight deep-learning perception and real-time inverse kinematic control are innovatively integrated. The perception layer employs the MediaPipe Lite convolutional network to detect 33 key human-body points at 30 fps.It also concurrently fuses the spatial coordinates from the depth camera, generating accurate 3D joint data. The mapping layer initiates a dynamic shoulder-reference calibration mechanism. This overcomes the traditional reliance on calibration and combines dual constraints of the working space, ensuring motion safety. The control layer features a geometric closed inverse-solution model (real-time solution in 0.8 ms). Through a two-threaded asynchronous architecture, it separates attitude detection from joint control, thoroughly resolving the bottleneck of response delay. Experimental validation indicates that on the JAKA Zu3 platform, the system attains a reduced terminal-trajectory tracking error, an extremely low action-delay rate, and a zero joint-overtravel rate. It suits the dynamic scenario of operators with heights ranging from 160 to 190 cm.In the future, it can find applications in settings like automotive assembly lines and nuclear-waste treatment, offering a highly adaptable human-machine collaboration model for flexible manufacturing.