Motion control of non-holonomic constrained mobile robot using deep reinforcement learning

For the motion control problem of non-holonomic constrained mobile robots, a point stabilization kinematic control law for mobile robot based on deep reinforcement learning is proposed. Firstly, a kinematic model of mobile robot is constructed to build memory for deep reinforcement learning, including the current state of the robot, the control action, the reward and the next state of the robot, which is generated through the connection between mobile robot and environment. Then, value network parameters in the real-time network are updated by a loss function, which is composed of a state-action value in current moment came from the value network of real-time network and a target value, the state-action value of next moment generated by the value network in target network. Next, the parameters of policy network of real-time network are updated according to the state-action value generated by value network of the real-time network in current moment. Finally, the parameters in the real-time network are weighted and averaged with the parameters in the target network, so the parameters of target network are updated to control mobile robot to stabilize with desired point. The simulation and experiment results show that the control algorithm based on deep reinforcement learning could effectively realize the point stabilization control of nonholonomic mobile robots.