Introducing Self-Learning into Robotic Arm Using Deep Reinforcement Learning
With the growing trend of autonomous machines, the combination of supervised and unsupervised machine learning techniques has been explored in providing optimal solutions for self-learning. In robotics, the curse of dimensionality makes convergence of machine learning difficult, no matter whether it is supervised or unsupervised. Therefore, reinforcement learning, which often requires a large number of trials for effective learning experience similar to unsupervised learning, suffers serious challenges in robotic applications. Consequently, choosing an appropriate algorithm that would perform optimally is of utmost importance. In this work, a robotic arm having 6 degrees of freedom combines supervised and unsupervised learning techniques by using a concept called Deep Reinforcement Learning, this helps the robot in becoming autonomous. It uses a camera image as an input to generate states through observation of the image, and distance for the reward system. It learns the optimum policy for action selection given a particular state observation that would achieve the maximum reward. The off-policy Deep Q Network (DQN) algorithm is to be implemented in this design and will be deployed on the robotic arm for independently learning the optimum movement towards achieving a certain task in a controlled environment.