Navigation System in Dynamic Indoor Environment using Deep Reinforcement Learning with Actor-Critic

The challenge of mobile robot navigation system is how to find the optimal path and be able to avoid obstacles. The obstacles could be stationary or moving objects, which make the environment of mobile robot would be changed dynamically. Regarding this problem, the conventional method should update the path everytime when there are any changes in environment, so it causes the computational burden. This paper aims to solve the problem by developing Deep Reinforcement Learning with Actor-Critic (DRL-AC) as navigation method to find the optimal path. DRL-AC determines robot movement by selecting action which have highest reward. DRL-AC is developed with actor and critic neural networks which are trained in dynamic environment. To prevent hardware damages during the training process, DRL-AC is developed in simulator. This simulator is designed to adopt real world condition. After training, DRL-AC can be implemented to the real mobile robot. The performance of DRL-AC is analyzed through neuron dimension and how the mobile robot navigated successfully in dynamic environment. The training results show that neural network with 2048 x 1024 hidden layer started to be convergent and mobile robot navigated successfully to avoid obstacle.