Abstract:
This research proposed model-based reinforcement learning (MBRL) algorithm for robotic arm control with 3D camera object detection for object manipulation. 3D camera was used to detect gripper, object, and target position in a 3D coordinates system as training data of the MBRL. Within MBRL, the covariance matrix adaptation evolution strategy (CMA-ES) was combined with machine learning techniques to create an environment model. The test problems are divided into numerical studies and real problems. The numerical studies were investigated with additional uniform noise in movement. The inverse kinematics (IK) was compared with 3 machine learning regression techniques used in the MBRL, Gaussian process regression (GPR), artificial neural network (ANN), and support vector regression (SVR). The results show that the GPR technique has the highest success rate of all numerical studies with a value as high as 100%, because GPR is approximating covariance method that considers noise. Although GPR spent the most training time, GPR was more suitable than other techniques of which the approximately average success rate was only 50%. Therefore, the MBRL(GPR) was used to create Actor network (AN) or MBRL(GPR)+AN. The results of the comparison between MBRL(GPR) and MBRL(GPR)+AN show that MBRL(GPR) still has the highest success rate of 100% and spent training time less than MBRL(GPR)+AN. In real problems, when MBRL(GPR) is combined with 3D camera object detection, The results show that MBRL(GPR) still has the highest success rate of 100% for both planar motion control and three-dimensional space. It can also pick and place object with 100% success.