Research On Reinforcement Learning Based Control Method Of Magnetic Navigation AGV

With the continuous development of the flexible manufacturing systems,automated warehouse system and other advanced production system,AGV(Automatic Guided Vehicle),as a wheeled mobile robot,plays an increasingly important role in logistics and advanced manufacturing systems.The path following control is a key technology to achieve AGV high precision guidance control.In order to avoid a significant amount of manual parameter tuning for the design of the magnetic navigation AGV path following controller when the system parameters are unknown and apply the previously learned knowledge to speed up training when the algorithm is applied to another AGV with different parameters,this thesis proposed two novel methods.The main content includes the following two aspects:(1)This thesis designs a path following controller using an improved model-free deep reinforcement learning algorithm.Based on the kinematic and dynamics model of magnetic navigation AGV derived in the discrete-time domain,the path following problem is formulated as continuous-state,continuous-action Markov decision process.Two neural networks are used to implement an Actor-Critic architecture off-policy model-free reinforcement learning algorithm as a controller implementation.The temporal-difference algorithm and the deterministic policy gradient algorithm are respectively used to update the parameters of Critic and Actor network to approach the optimal action value function and strategy function respectively.(2)Introduce representation learning method to speed up the training process when algorithm is applied to a new AGV with different parameters.By using auto-encoder technique to learn a general mapping from input signals to state representations to disentangles the state representation learning and the task-related strategy learning.When algorithm is applied to a new AGV,the pre-trained encoder replaces the first hidden layer of the network to implement the process of encoding the input signal into a state vector,thereby shortening the training process of the algorithm.Besides,this thesis presents a synchronous training auto-encoder method.The auto-encoder is trained synchronously by using the data generated during the training process of the tracking control algorithm,so that the training process of the auto-encoder is accelerated and the utilization rate of the algorithm is improved.