This paper presents a novel 2 DOF robotic arm wrestling system integrated with mechanical arm, elbow/wrist force sensors, servo motor, encoder, 3-D MEMS accelerometer, and USB camera. The arm wrestling robot (AWR) is used to play arm wrestling game with human for entertainment. Based on the force testing equipment, we acquire the data of surface electromyographic (EMG) signals form target muscle when a real player competes with the robot. Wavelet transform and neural network are applied to extract the characteristics of EMG signals and estimate the joint torque. Experiment results have proved the validation of the wavelet neural network method.