Reinforcement learning based algorithm design for robot manipulator task planning

• Main Author: Gui, Shun
• Other Authors: Hu, Guoqiang
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation
• Online Access: https://hdl.handle.net/10356/151199
• Institution: Nanyang Technological University

Artificial intelligence and machine learning as the most advanced, cutting-edge technologies have become a current research hotspot, bringing new research to many fields. Deep reinforcement learning can be used in robotics, which has great potential to enhance the intelligence of robots. Firstly, we performed three deep reinforcement learning algorithms, i.e. DQN, PPO, and A2C, into robot grasping learning and verified the feasibility of the algorithms. In order to perform the algorithm to continuous state and action scenarios, we adapted the algorithm to deal with the robot continuous control problem. Through simulation training, we obtained grasping manipulation policies that can be used for robot grasping. Unlike traditional grasping solutions, we also defined robot grasping scenarios based on a framework of reinforcement learning, including state, action, reward, multi-step Markov decision process, etc. Based on this framework, the grasping manipulation was fully transformed into a reinforcement learning problem and the deep learning algorithm can be completely performed on it. The final robot achieved the learning of grasping skills. Furthermore, we built a complete deep reinforcement learning robot grasping simulation program based on some software. All simulations were executed in ROS in Ubuntu system. PR2 was designated as the agent, who implemented the learning process and grasping manipulations. Gazebo was used to simulate a physical environment, where the robot could execute a grasping manipulation like in the real world. MoveIt was used to do motion planning after the agent chose an action at current state. Also, other manipulations can be simulated in this program. Finally, we performed some grasp tests on the obtained grasping policies and compared the performance of three reinforcement learning algorithms in grasping learning. To enhancing the performance of the algorithm, we will consider including dynamic parameters to the reinforcement learning process to improve the control of the robot’s velocity, acceleration and contact force during grasping in the future. Since robots need to obtain the location of objects in advance during the learning process, we will also make the robot calculate the object location from sensors instead of the current manually given location data.