During drilling, a rotary steerable system (RSS) is affected by vibration, rotation, and other random noises. This paper presents a random weighting adaptive estimation of model errors on attitude measurement for RSS drilling tools. The algorithm is used to estimate the covariance matrix of the dynamic model errors and the state prediction vectors for RSS tools, to reduce the effect of abnormal noises of the dynamic model on the estimation of state parameters. The random weighting estimations of observation's systematic error and covariance matrices of observation error vectors to control the effect of observation model noises anomaly on state parameter estimation, enhance the use efficiency of the latest observation information, indirectly weaken the impact of model error on state parameter vector, and improve the accuracy of attitude calculation of RSS. Experimental results and comparison analysis demonstrate that the proposed algorithm achieves better outcomes than the Kalman filtering and Extended Kalman filtering. The maximum solution error is controlled at 0.15°, and the tool face angle error is less than 3°. It cannot only control the covariance matrix of the observation error vector and predicted residual vector, but it can also effectively resist the disturbances of kinematic model error. It also can improve the dynamic measurement accuracy of attitude parameters, and solves the problem of uncertainty in dynamic attitude measurement of rotary steerable drilling tools.