| With the continuous innovation of science and technology,the field of artificial intelligence has developed rapidly.A range of products represented by intelligent robots and AI technology are gradually integrated into the lives of many consumers.Intelligent robotics has been commonly developed in the logistics and warehousing industry,but the "last mile" of delivery in the courier industry is mainly done manually,which is not only a large workload but also consumes a lot of human and material resources.In this project,the key technologies of logistics delivery vehicle with intelligent addressing and obstacle avoidance is studied to design an intelligent logistics delivery vehicle that can replace manual work for efficient and fast delivery.In order to achieve the above objectives,we carried out the research in three aspects:physical structure,positioning algorithm and path planning,and the main research contents are shown as follows:(1)Overall design of the logistics vehicle.The advantages and disadvantages of current mainstream vehicle structures were analyzed,and the chassis model with Ackermann steering mechanism as the core was selected.The two-dimensional and three-dimensional dimensional drawings of the chassis system were designed,and the physical chassis was built.And the kinematic analysis of the logistics vehicle was carried out to convert the twoaxis speed into the motor speed and steering angle of the wheels.The master-slave control scheme with ARM processor as the core and the related hardware selection were determined.The chassis software system of the lower computer was designed,and the communication method and data structure between the ROS system of host computer and the STM32 of lower computer were defined.(2)Optimized the map building accuracy of SLAM algorithm.Analyzed the advantages and disadvantages of current open source algorithms and selected the algorithm framework.The multi-sensor fusion method of combined IMU data was designed for the low accuracy and poor robustness of single Li DAR map building.The positional information obtained from IMU integration was used as the initial value of LIDAR point cloud matching to improve the accuracy of positional calculation,and finally the positional information from two sensors was weighted and fused to improve the accuracy of map building.(3)Improved the path planning algorithm.For the traditional path planning needs to build a global map,which is time-consuming and has poor adaptability to the environment,an intelligent path planning algorithm based on deep reinforcement learning was used.Reinforcement learning was used as a model to fit the return values of different states using neural networks.Traditional deep reinforcement learning has the problem of low efficiency and not easy to converge.In this paper,we proposed an improved deep reinforcement learning method by setting a gradient reward policy,which made the reward of environmental feedback continuous and negatively correlated with the distance to the end point,and optimizes the exploration strategy to achieve the best exploration of the environment.After the improved algorithm proposed in this paper was simulated several times and physically demonstrated,the results showed that the multi-sensor fusion method combining IMU and LIDAR can reduce the error of single LIDAR data by 0.28 m on average,optimized the quality and accuracy of SLAM map building,and improved the robustness.The path planning algorithm based on improved deep reinforcement learning could improve about50% in efficiency,while alleviated the problem of easily falling into local optimum,and stabilized the path planning success rate above 80% in the late exploration period.During the outdoor road test,the vehicle could effectively perform obstacle avoidance driving,which confirmed the practicality of the improved algorithm in this paper. |
PDF Full Text Request