The majority of today`s industrial robots are programmed to follow a predefined trajectory. This is sufficient when the robot is working in a fixed environment where all objects of interest are situated in a predetermined position relative to the robot base. However, if the robot`s position is altered all the trajectories have to be reprogrammed for the robot to be able to perform its tasks. Another option is teleoperation, where a human operator conducts all the movements during the operation in master-slave architecture. Since any positioning errors can be visually compensated by the human operator, this configuration does not demand that the robot has a high absolute accuracy. However, the drawback is the low speed and low accuracy of the human operator scheme. The manipulator considered in this thesis is attached to a ROV (Remote Operating Vehicle) and is brought to its working environment by the ROV operator. Every time the robot is repositioned, it needs to estimate its position and orientation relative to the work environment. The ROV operates at great depths and there are few sensors which can operate at extreme depths. This is the incentive for the use of computer vision to estimate the relative position of the manipulator. Through cameras the differences between the actual and desired position of the manipulators is estimated. This information is sent to controllers to correct the pre-programmed trajectories. The manipulator movement commands are programmed off-line by a CAD system, without need even to turn on the robot, allowing for greatest speed on its validation, as well as problem solving. This work includes camera calibration and calibration of the structure of the manipulator. The increased accuracies achieved by these steps are merged to achieve in-situ calibration of the manipulator base.