Deep Reinforcement Learning Algorithm for Object Placement Tasks with Manipulator

To settle the problem that the household robot needs to have the flexibility of object types and target poses when performing object placement tasks, a deep reinforcement learning algorithm is utilized in this paper. By using this algorithm robot can learn the placement method for different object types autonomously under the policy search part, and a convolutional neural network (CNN) policy is trained to make the robot adapt to different target poses. When performing these tasks, the placement error can also be reduced by modifying the sampling method and the weight form of cost function. Finally, the learning ability and flexibility of the deep reinforcement learning algorithm is tested by simulation.