A learning based approach to self modeling robots

Performance of a robotic system depends o n the mathematical model that is programmed within the software of the robot. Usually the mathematical model of a robot is calculated by the designer and is re-calibrated during the time of operation. Model of the robot is essential for its software to take intelligent actions. Equipping the robot with the capability to self model will help itself to re-correct the model and recover itself in case of minor damages during operation. This also avoids the need of re-calibration since the model gets refined with more number of iterations. This work tries to make a robotic system, where the robot attempts to model itself by making use of some basic prior information and deriving the rest of the required information by visually observing the result of its actions on the environment. The approach discussed in this paper is for robotic manipulators and is validated through an experiment using an Invenscience ARM 2.0 using a Point Grey stereo camera.