Self-Supervised Learning of Robot Manipulation

Robot arm manipulation with reinforcement learning is an active research area. Immitation learning from human is one of the most curious till date to researchers. Demonstration video based learning becomes a challenge when the view angle becomes an independent parameter. Using third person view of some similar task and robot’s third person view, many complex actions can be trained using time coherence property of similar features from TCN (Time Contrastive Network). Learning and controlling a manipulator using its dynamics very difficult and time consuming process. So in this paper we are using kinematic model based PID control for model based arm control. Model free control is implemented based on novel Reward based Random Exploration (RRE) algorithm. This model based and model free updates are memorized and uses it to complete the task.