A Reinforcement Learning Approach for Inverse Kinematics of Arm Robot

The inverse kinematics is the foundation and emphases of the industrial robot control. Traditional solutions of inverse kinematics cause many difficulties to the exploitation of many kinds of industrial robots because of the complexity derivation, difficulty of calculation, multiple solutions, lack of instantaneity. This paper proposes a new way to obtain the inverse kinematics of 5-DOF arm robot with a grip by using the method of deep deterministic policy gradient in reinforcement learning, the method combines the neural network and robotics knowledge through the continuing attempts to get the accuracy solution. The propose of the simulation by Tensorflow and Matplotlib is designed to verify the accuracy of the new way, the results of simulations show that comparing with the traditional way, the end grip joint of robot can arrive at the location we set with some more error, but the angle of every joint can be calculated and the error is in an acceptable range, the accuracy of the angle and posture is satisfied. This is a new way to solve inverse kinematics of robot which is easier than traditional way, but has more meaning on solutions of multi-degree of freedom robots.