Abstract—In this paper, an approach based on multilayer neural network is proposed for human-robot collisions detection. The neural network is trained by Levenberg-Marquardt algorithm to the dynamics of the robot with and without external contacts to detect unwanted collisions of the human operator with the robot using only the proprietary position and joint torque sensors of the manipulator. The proposed method is evaluated experimentally using the 7-DOF KUKA LWR manipulator and the results illustrate that the developed system is efficient and very fast in detecting the collisions.