Robotic Crowd Navigation Based On Deep Reinforcement Learning With Interactive Attention

In recent years,with the rapid development of robotics,mobile robots have gradually integrated into the daily life of human beings,and dynamically dense pedestrians in the crowded environment make it difficult for robots to navigate.How to make the robot plan a collision-free path efficiently under the social norms of human walking is a research hotspot in the field of mobile robots,which has important practical significance and research value.But making a mobile robot find an efficient path in a short period of time usually requires predicting the robot’s interactions with neighboring pedestrians.As crowds grow and scenarios become more complex,researchers often reduce the problem to a one-way human-computer interaction problem.But in fact,we should not only consider the interaction between human and robot,but also consider the impact of human-human interaction on the trajectory of the robot.Therefore,this paper proposes two methods for robot navigating in crowd based on deep reinforcement learning of interactive attention,so that the robot can avoid obstacles in the crowd and smoothly navigate from the starting point to the target point.The first model first combines the state of the robot,the state of the human,and the characteristics of human-human interaction to obtain the joint state that the mobile robot perceives in the current environment after a series of processing.And according to the principle of attention mechanism,a dual attention module that can adaptively weighted and aggregate the important feature data of the joint state is designed to model the interaction between robots and people and between people,so that the robot can give different attention weights to pedestrians in different states.Finally,the reinforcement learning algorithm is used to train the model,and social norms are included in the design of the reward function,so that the robot’s decision-making is more social.To verify the validity of the model,a simulation experiment environment is built.The model is trained in this environment,and the model is compared with the current representative navigation algorithm in the crowd environment.It has better performance in terms of time and sociality,and enables the robot to navigate efficiently and follow social norms in crowded environments,which verifies the effectiveness of the method.The second model analyzes the robot navigation graph and models the interaction features through relational reasoning.Then the model uses a graph convolutional network to calculate the paired interaction features.Next,it uses the attention mechanism to extract the important features in the interaction features,and estimate the value of the joint state and predict the human motion,respectively.Finally,the model is placed in the same simulated experimental environment as the first model for training and testing.Compared with the current more advanced methods,the obtained experimental data shows that the model performs better.In future research,it is considered to further extend the algorithm so that the robot can output continuous actions and complete the navigation more efficiently and safely.