Reinforcement Learning assisted LQR Tuning of a Load Bearing Robotic Arm

Abstract In this paper, a Linear Quadratic Regulator feedback controller that is tuned by a Reinforcement Learning agent is implemented into the control of a DOBOT Magician Robot Arm to carry and place a load, to which the performance will be compared to the same action undergone by the default movement control of the DOBOT Magician. A 3D URDF model of the DOBOT Magician is imported into Simulink in MATLAB to derive a state-space model. The state-space model is then shaped in Python together with an LQR controller. The Q and R values, which correspond to the state deviation and control effort of the LQR controller, are iteratively tuned by a Reinforcement Learning agent created through TensorFlow in Python, being trained to maximise the efficiency of a particular movement that is performed under a set time frame. Once the Q and R values have been set, the movement efficiency is then benchmarked against the efficiency of the default DOBOT movement.