The Determination of Reward Function in AGV Motion Control Based on DQN

Abstract

Motion control is a very important part in the field of AGV(Automated Guided Vehicle). A good motion control method can make the movement of AGV more stable. Network models of reinforcement learning is one of the methods to solve the problem of AGV in motion control. This paper introduces the Markov Decision Process and the role of reward function. Besides, it studies and analyzes several classic reinforcement learning cases. DQN(Deep Q-Learning Network) which belongs to deep reinforcement learning network model is adopted. We set up several sets of comparative experiments with different reward functions by using sparse reward setting method, formalized reward setting method and reward coefficient variation reward setting method. Also we adjust the time of training. Through comparison, a reward function suitable for solving the problem of AGV motion control is obtained. In the field of AGV motion control, reinforcement learning model can converge faster and make more correct decisions. The reward function is verified in the simulation environment built by ROS(Robot Operating System).

Autonomous Navigation of Robots: Optimization with DQN

“…The robot can be trained by interacting with its environment and receiving feedback in the form of rewards or penalties based on the actions it takes. For example, the robot may receive a reward for successfully navigating to a particular location, while receiving a penalty for colliding with an obstacle [33][34][35][36][37]. Over time, the robot can use the feedback it receives to learn the optimal path to take in different environments.…”

Research on the Local Path Planning for Mobile Robots based on PRO-Dueling Deep Q-Network (DQN) Algorithm

“…All the parameters in the training network are assigned to the target network after the training of a fixed number of steps. The algorithm sets an experience replay unit to reduce the correlation of training samples and improve the instability of the action value function of neural network approximation reinforcement learning [19]. A batch of samples are evenly selected from the experience library and mixed together with the training samples to break the correlation between adjacent training samples and improve the utilization rate of the samples during each training.…”