Long-Term Autonomous Navigation without HD Maps

In this research, we build an autonomy system for biped robots to navigate outdoors and indoors to multiple destinations with a known topological map. The autonomous system includes: (1) topological mapping to generate topological map indoors; (2) a global route planner to generate the path in the topological map for navigation; (3) state estimator to estimate odometry poses of robots; (4) feature extraction to detect features in the environment for localization and planning; (5) localization to find the robot pose in the topological map; (6) reactive planning system to generate the control commands for robots. The indoor topological maps are generated by calculating the Voronoi graph of the occupancy grid map of the experiment environment. There are two options for the global route planner: 1) Informable Multi-Objective and Multi-Destinations RRT* (IMOMD-RRT*) that finds the shortest path visiting all the destinations; 2) Picture to OSM which detects the hand-drawn path on the printed map. Contact-aided invariant extended Kalman filtering state estimator is used to estimate robots' odometry pose by using IMU and contact sensor dynamics. For outdoor experiments, the curb features are extracted from LiDAR point clouds. Moreover, semantic point clouds are generated by LiDAR point clouds and camera images by using Support Vector Machine (SVM) and dynamic programming. For indoor experiments, wall features are extracted from LiDAR point clouds. Localization is implemented by using particle filter to match the extracted features and the landmarks in the topological map. The elevation map and terrain map are built by the LiDAR point clouds and robot poses. The reactive planning system chooses the intermediate goals and generates the optimal control commands to the goals.