In order to make the arm rehabilitation patients accurately feel the endofarm force, a soft measurement method of endofarm force is proposed based on random forest regression. The data from the electromyography (EMG) sensors and the arm gesture sensors are integrated to describe the comprehensive feature information of the arm. The comprehensive feature information is used as the input of Random Forest Regression with the endofarm force as the output. According to the basic actions of arm rehabilitation, two groups of experiments ‘pushpull’ and ‘liftdrop’ are specifically designed. The Random Forest Regression is trained offline with the comprehensive feature information and the endofarm force measured with a force sensor, then prediction of endofarm force is conducted and compared with the real force. The method is verified by the experimental results and the effectiveness of the method is proved.