Research On Perception And Navigation Methods Of Service Robots For Crowd Scenes

Autonomous navigation is the most fundamental and important function of mobile service robots.Traditional navigation methods typically achieve goal-oriented navigation in deterministic environments through manually-designed rules.However,in the context of human-robot interaction,the high dynamics and social nature of pedestrian movement pose new challenges for autonomous navigation of service robots.This paper focuses on developing perception and navigation methods for service robots in complex crowded environment,using deep reinforcement learning as the primary technical means,with the goal of achieving safe,comfortable and socially compliant human-like navigation.The main innovative contributions of this this work are as follows:In response to the lack of unified safety standards for service robots in crowded environments,this paper proposes a method for judging the action risk of service robots based on pedestrian social psychology.From the perspective of social psychology,a personal space theory model is introduced to determine the safe distance of the robot in the position,and combined with the robot-pedestrian relative motion speed and human reaction time,the trust speed condition of the robot is proposed.In order to solve the problem that existing methods lack effective safety mechanisms to deal with fast-moving pedestrians,this paper proposes a design of reinforcement learning safety reward function,and models and analyzes the robot-pedestrian collision process using the impulse-momentum model,using the damage caused by the collision impulse reaction as the quantitative index of the action risk of the robot under different speed conditions.Simulation and experimental results in real environments show that compared with the current optimal method,the proposed method reduces the average navigation collision rate by 6 percentage points and the average risk action rate by 35.7%,demonstrating higher levels of safety.In real environments,service robots need to simultaneously perceive and understand the interactive behavior of the crowd and other obstacle information.In response to the shortcomings of existing models in pedestrian importance evaluation,this paper designs a pedestrian importance evaluation network that combines static and dynamic object features.The network adopts a social attention mechanism to integrate static obstacle and dynamic pedestrian information,and establishes a more optimized pedestrian importance evaluation structure,improving the ability of the model to process pedestrian interactive features.Meanwhile,this paper proposes a hybrid navigation method to address the lack of robot kinematic constraints and effective handling mechanism for static obstacles in existing models.The method redesigns the reinforcement learning action space and combines the rule-based trajectory generator into the deep reinforcement learning method.Reinforcement learning outputs local goal intentions to the multi-constraint trajectory generator in the back-end to complete robot control,instead of directly outputting robot control commands.Simulation comparison experiments and module ablation experiments have verified the effectiveness of the proposed perception and navigation algorithm in handling comprehensive obstacle information.Among them,compared with the current optimal comparison method,the proposed method improves the average navigation success rate by 25 percentage points,enhancing the navigation effect of the model in complex social scenarios.