Research On Interactive Indoor Robot Navigation Algorithm Based On Reinforcement Learning

At present,robot technology has been widely used in various fields.Compared to the widespread application of robots in industrial production,the development of server based robots in daily life has only just begun.One of the key issues that need to be addressed in the landing application of service-oriented robots is robot navigation.Currently,for some simple environments,robots can complete navigation tasks through some traditional navigation algorithms or learning based navigation methods efficiently.Therefore,the demand for robot navigation algorithms in complex environments is also increasing.Regardless of traditional navigation algorithms or learning based navigation methods,robot navigation in these environments with a large number of mobile pedestrians and complex obstacles should be vigorously developed.Therefore,according to the scientific research needs during the experimental process,this thesis will conduct research on navigation algorithms in complex pedestrian environments based on reinforcement learning.For such complex environments,dense crowds can easily significantly affect the performance of robots and even cause them to collide.Therefore,many robot navigation algorithms are dedicated to making robots strive to avoid pedestrians and ensure safety.However,in many narrow or crowded pedestrian scenes,robots will have to come into close contact with humans.In this case,robots may not be able to find a path that can advance based on their own abilities,and it is also dangerous to forcibly navigate through gaps in a flowing crowd.If the robot can actively interact with pedestrians,such as making a sound to remind pedestrians to avoid collision,it can greatly improve the efficiency and safety of navigation.Such an algorithm is also known as an interactive robot navigation algorithm.However,both learning based and traditional methods often face many challenges when designing interactive navigation algorithms.Especially for reinforcement learning,although using reinforcement learning can facilitate and quickly train robots,relying solely on sparse rewards may be difficult for robot t achieve good training results.Therefore,this thesis will conduct research on interactive robot navigation algorithms based on reinforcement learning.This thesis will design an algorithm that can effectively train robots to navigate interactively in dense environments based on the PPO algorithm in reinforcement learning.By building a simulation platform to simulate the real environment,robots can simulate sensing data from the platform environment,and rely solely on the output of neural networks to determine their behavior in different surrounding environments.Under this algorithm,the robot can demonstrate pedestrian friendly,intelligent,generalized,and efficient navigation strategies in various complex environments through simple training.Through rational use of human interaction,robots can complete navigation tasks with a higher arrival rate than other algorithms.For reinforcement learning navigation,transitioning from a simulation platform to real-world applications is another difficulty.The real environment faced by robots often has a large gap from the environment for intensive learning and training,resulting in the robot being unable to adapt to the real environment.Due to the large variation in the output of the neural network,the shaking of the robot during the navigation is very dangerous.Through our research and analysis,we have identified relevant factors that may lead to poor robustness and stability of reinforcement learning navigation,and propose corresponding methods to address them in our subsequent work.For example,for the problem of poor robustness,we have established more complex and diverse pedestrian models to retrain the robot.For the problem of poor stability,we will propose a new reinforcement learning reward function to optimize its stability and smooth the speed of the robot during operation.Through the above work,we have verified that the reinforcement learning navigation algorithm can significantly overcome the above problems through our series of optimizations.Finally,we applied the interactive navigation algorithm we proposed to a real-world robot.By combining various modules of the robot with the reinforcement learning navigation algorithm,we successfully applied it to a real environment,verifying the feasibility of the algorithm.