Abstract: A sensorimotor joint learning approach is proposed to address the problems that the learning process of autonomous robots may be limited by the input data during training, and the perceptual space and the motor space are separated. The proposed approach establishes a sensorimotor system through the human simulated intelligent schema theory, divides the system into perceptual, associative and motor schemas, and establishes the association between perceptual and motor schemas based on active inference. For the robot's active exploration and learning process, an associative model and a cognitive module are designed, and a joint learning algorithm is proposed to combine active perception for the robot self-learning. The experimental results show that the joint learning approach can improve the perception accuracy by 22.32% in simulation and 12.00% in the physical system, and significantly reduce the time to reach the same target position, and the movement trajectory becomes smoother. The experiments demonstrate that the proposed joint learning method can enhance the robot's cognition and self-learning abilities, as well as can autonomously optimize recognition accuracy and movement trajectory.