An improved gradient descent algorithm based on threeaxis gyroscope, threeaxis accelerometer and threeaxis magnetometer was proposed to solve the problem of low accuracy of attitude estimation during mobile robots movement. The improved gradient descent method had the advantages of less computation and higher accuracy. It accelerated the gradient descent process by introducing momentum, which improved the convergence speed of the original algorithm.By introducing the accelerometer confidence, the error caused by the drastic change of accelerometer caused by external factors could be effectively reduced.The experimental results show that compared with the traditional algorithm in the case of the accelerometer with or without interference, the accuracy of the improved algorithm is improved by about 3°and 0.5°respectively