Research On Navigation Algorithms Of Mobile Robot Based On Control Variable Optimization And End-to-end Learning

With the development of robot technology,mobile robots have been applied in various fields such as disaster relief,industrial production,and social services.In these areas,mobile robots need to move autonomously and quickly through a complex dynamic environment.Autonomous navigation algorithm is the key technology of intelligent mobile robots.This thesis focuses on the non-holonomoic mobile robots including differentialdrive robots and car-like robots,and conducts research on navigation problems with and without a pre-built map.With a pre-built map,this thesis proposes a time-optimal local planning algorithm for mobile robots based on optimizing control variables of differential-drive and car-like mobile robots.The optimization-based planning model of differential mobile robots and car-shaped mobile robots are established.Analytical gradient expressions of optimization functions for objective variables and constraint functions are derived.A new type of projection gradient descent and conjugate gradient descent optimization method is introduced to achieve fast solution of proposed local planning algorithm.The optimization problem is solved in the feasible control space of the robot,and it is more direct to deal with kino-dynamincs constraints and non-holonomic constraints.The performance of the navigation algorithm,such as time optimality,real-time performance,robustness,and control continuity,was verified using simulation and experiments.In the absence of obstacles,we verified that the trajectory planned by the proposed method is very close to the theoretically time-optimal Reeds and Shepps trajectory.And the planned trajectory is better than the Fraichard trajectory of continuous curvature.In order to deal with the situation where there are multiple homotopy classes from the starting pose to the goal,a time-optimal local planning algorithm for mobile robots based on multiple homotopy classes exploration is proposed.First,the problem of uni-homotopy in local planning is analyzed.Later,H tags were introduced to distinguish multiple homotopy classes,and a sparse probability map was used to quickly obtain the topological information of the environment.With the help of homotopy classes,multiple initial values of optimized trajectories can be introduced to avoid local optimization of local optimization issues.Compared with existing methods using Voronoi diagrams and dense probability maps,our method is characterized by maintaining multiple homotopy classes using sparse probability maps,and it is faster in searching for distinctive homotopy classes.The simulation results show that the navigation algorithm can handle the optimal path planning problem of multiple homotopy cases and can find the global optimal solution.In order to solve the problem that the robot has only low-cost sparse lidar navigation without a pre-built map,a map-less navigation method based on end-to-end learning is proposed.First,a neural network controller is proposed,and the training framework of the neural network controller is designed.By imitating an expert planning strategy,the neural network planner without a pre-built map is learned.The results show that the mapless planning algorithm can complete the navigation task in complicated situations,and the navigation path is not much different from the map-based navigation algorithm,but the navigation time is relatively high.