In order to compensate for the end effector attitude error caused by insufficient torsion of the robot joint and the end effector connection, a method based on inertial measurement unit to obtain the end attitude online is proposed. First of all, the motion process of the entire robotic arm system is divided into static and dynamic processes. At static, due to the small external acceleration noise, a method for estimating the attitude angle of the end effector based on local gravity using an accelerometer is proposed. In dynamic time, an adaptive extended Kalman filtering algorithm based on noise statistics is proposed for the problems of external acceleration noise, gyroscope zero drift, and scale factor error that affect the measurement accuracy. Based on the measurements of the accelerometer, the weights of the observed noise variance array are updated to adjust the Kalman gain and reduce the effect of acceleration noise on the measurement accuracy. Experimental results show that the average attitude angle error estimated by the static algorithm is 0.07°, 0.05°, 0.2°; In dynamic time, the proposed algorithm can compensate well for the influence of external acceleration on attitude, and can effectively improve the attitude measurement accuracy, compared with the EKF algorithm, the average error of attitude angle is reduced by 2.69°, 1.01°, 0.5°.