Redundant Robot Control Using Multi Agent Reinforcement Learning

Robot control in task-space1 needs the inverse kinematics and Jacobian matrix. They are not available for redundant robots, because there are so many degrees-of-freedom (DOF). Intelligent learning methods, such as neural networks (NN) and reinforcement learning (RL) can learn them. However, NN needs big data and RL is not suitable for multilink robots as the redundant robots. In this paper, we propose a full cooperative multi-agent reinforcement learning (MARL) to solve the above problems. Each joint of the robot is regarded as one agent. Although the dimension of the learning space is very large, the full cooperative MARL uses the kinematic learning and avoids the function approximators in large learning space. The experimental results show that our MARL is much more better compared with the classic methods such as, Jacobian-based methods and neural networks.1Task-space (or Cartesian space) is defined by the position and orientation of the end effector of a robot. Joint-space is defined by angular displacements of each joint of a robot.