The invention relates to the field of self-adaptive control, in particular to a mechanical arm kinematics self-calibration method based on binocular vision. Firstly, data are collected, and the tail end position information of a mechanical arm and the actual rotation angle and the theoretical rotation angle of each joint of the mechanical arm are obtained through a binocular camera. Then, the DH theory and the hand-eye calibration theory are fused, and a mechanical arm kinematics hybrid model is established. The model is trained by using a multi-population self-adaptive difference algorithm, and parameters of the hybrid model are solved. Finally, each servo motor model is established through a polynomial fitting method, and polynomial parameter solving and compensation prediction are conducted by using a least square method. According to the hybrid model and the servo motor models provided by the invention, the influence of geometric errors on the mechanical arm can be greatly reduced, and more practical model parameters can be calculated. A mechanical arm base coordinate system needed in the hand-eye calibration process does not need to be additionally established through a demonstrator, and automation of the whole calibration process can be achieved on the premise that absolute positioning precision is guaranteed.