Task-space control needs the inverse kinematics solution or Jacobian matrix for the transformation from task space to joint space. However, they are not always available for redundant robots because there are more joint degrees-of-freedom than Cartesian degrees-of-freedom. Intelligent learning methods, such as neural networks (NN) and reinforcement learning (RL) can learn the inverse kinematics solution. However, NN needs big data and classical RL is not suitable for multi-link robots controlled in task space. In this paper, we propose a fully cooperative multi-agent reinforcement learning (MARL) to solve the kinematic problem of redundant robots. Each joint of the robot is regarded as one agent. The fully cooperative MARL uses a kinematic learning to avoid function approximators and large learning space. The convergence property of the proposed MARL is analyzed. The experimental results show that our MARL is much more better compared with the classic methods such as Jacobian-based methods and neural networks.

任务空间控制需要逆运动学解决方案或Jacobian矩阵，以实现从任务空间到关节空间的转换。但是，它们并不总是适用于冗余机器人，因为联合自由度比笛卡尔自由度更多。智能学习方法（例如神经网络（NN）和强化学习（RL））可以学习逆运动学解决方案。但是，NN需要大数据，传统的RL不适合在任务空间中控制的多链接机器人。在本文中，我们提出了一种完全协作的多智能体强化学习（MARL），以解决冗余机器人的运动学问题。机器人的每个关节都被视为一个代理。完全协作的MARL使用运动学学习来避免函数逼近器和较大的学习空间。分析了提出的MARL的收敛性。实验结果表明，与经典方法（如基于Jacobian的方法和神经网络）相比，我们的MARL更好。