Deep Q-Learning for Navigation of Robotic Arm for Tokamak Inspection

Computerized human-machine interfaces are used to control the manipulators and robots for inspection and maintenance activities in Tokamak. The activities embrace routine and critical activities such as tile inspection, dust cleaning, equipment handling and replacement tasks. Camera(s) is deployed on the robotic arm which moves inside the chamber to accomplish the inspection task. For navigating the robotic arm to the desired position, an inverse kinematic solution is required. Such closed-form inverse kinematic solutions become complex in the case of dexterous hyper-redundant robotic arms that have high degrees of freedom and can be used for inspections in narrow gaps. To develop real-time inverse kinematic solver for robots, a technique called Reinforcement Learning is used. There are various strategies to solve Reinforcement problem in polynomial time, one of them is Q-Learning. It can handle problems with stochastic transitions and rewards, without requiring adaption or probabilities of actions to be taken at a certain point. It is observed that Deep Q-Network successfully learned optimal policies from high dimension sensory inputs using Reinforcement Learning.