Research On Automatic Driving Decision Based On Responsibility Sensitive Safety Model

Currently,safe autonomous driving has become a future trend.For the decision-making of self-driving vehicles,the most important thing is to ensure the safety of the vehicle,and then other factors such as the efficiency of the vehicle need to be considered.The safety of the vehicle is the trustworthy foundation of the autonomous driving model.The inexplicability and lack of safety of deep reinforcement learning hinder its further application in the field of autonomous driving.How to build a safer autonomous driving model has become a key issue.Therefore,this thesis focuses on how to establish a safer model for deep reinforcement learning in single-vehicle and multi-vehicle collaboration scenarios.The specific content includes the following two aspects:Firstly,for reinforcement learning,there is no ability to predict danger,insufficient security and poor initial strategy,which requires meaningless trial and error.This thesis proposes a safety constraint method using the responsibility sensitive safety model,separates the potential dangerous action set and safety action set of the vehicle through the responsibility sensitive safety model,uses the deep reinforcement learning algorithm to make safer decisions under the supervision of the responsibility sensitive safety model,and realizes a safer automatic driving decision model,It reduces meaningless trial and error in the initial stage.Furthermore,the model is extended to multi vehicle system by using the method of parameter sharing,and the decision-making of multi intelligent vehicle is realized.Secondly,aiming at the problem that Multi-Agent Reinforcement Learning can not solve the dimension explosion of joint action space,this thesis adopts the multi-agent learning algorithm based on the average field theory,which equates the neighborhood vehicle of its own vehicle into a virtual vehicle through the average field theory,and transforms the interaction between its own vehicle and the neighborhood vehicle into the interaction between its own vehicle and the virtual vehicle,The problem of dimension explosion of joint action space is solved.Furthermore,using the responsibility sensitive safety model combined with the mean field multi-agent learning algorithm,the RSS security supervisor is used for security constraints on each vehicle in the system to improve the security and efficiency of the system.Finally,the proposed algorithm is tested and verified in the highway scene.The experimental results show that the deep reinforcement learning based on the responsibility sensitive safety model can make decisions under the security constraints of the model,and finally learn more secure and efficient strategies.In the large-scale multi vehicle cooperation scenario,the mean field multi vehicle cooperation algorithm simplifies the interaction between vehicles through the mean field,which can still have good results when other multi-agent algorithms can not be used.Furthermore,the responsibility sensitive safety model combined with the mean field multi vehicle cooperation algorithm improves the overall revenue and security of the system by improving the individual security and efficiency,and improves the security performance of the system in the scene of large-scale vehicle cooperation.