<p>Owing to the emergence of a need for robots that can handle heavy objects in fields such as logistics automation, a demand for high-payload collaborative robots has been increasing. The use of a counterbalance mechanism (CBM) in collaborative robots can significantly increase the payload without increasing the motor and speed reducer capacity. In this study, we propose a calibration method to reduce the error in the compensation torque generated by link-type CBMs that can be embedded inside links. The method involves the Levenberg–Marquardt algorithm and is based solely on the motions of a robot without the need to separate the CBM from the robot. To improve the calibration performance, the measured compensation torque was analyzed and appropriate calibration parameters were selected. The calibration results converged stably and the error in the compensation torque was significantly reduced. Consequently, the performance of robot control and external torque estimation has improved, thus resulting in improved hand-guiding and collision detection performance.</p>