Autonomous navigation method and system for intelligent indoor logistic transportation based on one-shot imitation

A navigation method for improving efficiency and safety of indoor logistic transportation based on an autonomous means of transportation, the method implemented by a computing system coupled to a LiDAR sensor and an odometry sensor, the method comprising:

• receiving, by the computing system, real-time geometric measurements from the LiDAR sensor and incremental pose measurements from the odometry sensor of the means of transportation;
• constructing, by the computing system, a navigable sparse graph of a three dimensional (3D) pose space based on an expert driving demonstration, the constructing comprising:
• (1) applying a method to the received measurements to generate a map of an indoor environment, the map indicating positions of static obstacles;
• (2) identifying, from the expert driving demonstration, safe positions of the means of transportation, each safe position being spaced apart from static obstacles by a predetermined safe distance;
• (3) treating the identified safe positions as a plurality of safe and achievable nodes in a discretized 3D pose space;
• (4) deriving a plurality of sparse 3D points in the 3D pose space as safe points based on linear interpolations of consecutive safe poses;
• and
• (5) saving only connected poses as the navigable sparse graph, thereby generating a costmap optimized for safe navigation within the physical constraints of the means of transportation in the indoor environment;
• generating, by the computing system, a robust localization estimation of the means of transportation's current pose within the generated map, the generating comprising:
• (1) fusing the real-time geometric measurements from the LiDAR sensor and the incremental pose measurements from the odometry sensor using a particle filter framework to obtain incremental poses;
• and
• (2) evaluating the obtained incremental poses based on a likelihood of point cloud measurements given a particle pose and the generated map, thereby providing accurate and real-time positional data for autonomous movement in dynamic indoor settings, even with odometry inaccuracies;
• determining, by the computing system, a global path from a current location to a target location within the indoor environment by applying a dynamic programming method to the navigable sparse graph, the dynamic programming method configured to efficiently re-compute the global plan in response to changes in the means of transportation's estimated position, thereby ensuring adaptable and continuous long-range planning based on learned safe routes;
• and
• executing, by the computing system, real-time local planning commands for directing the means of transportation along the global path while dynamically avoiding surrounding static or dynamic obstacles, the executing comprising:
• (1) randomly sampling a plurality of input sequences as a prediction of future actions to be taken by the means of transportation;
• (2) computing predicted positions and spatial boundaries of the means of transportation for each of the randomly sampled plurality of input sequences;
• (3) filtering out trajectories causing a collision between the means of transportation and the surrounding obstacles based on current LiDAR measurements;
• (4) computing a synthetic score for remaining trajectories, the score considering proximity to the global path and spacing from local obstacles;
• (5) determining whether an optimal trajectory exists among the remaining trajectories that maximizes the synthetic score for guiding the means of transportation and selecting the optimal trajectory if it exists;
• and
• (6) if no optimal trajectory exists, activating a recovery mode by adopting a clear-and-wait policy to wait for an obstacle to disappear or be removed while moving slightly to stay away from the obstacle, thereby enhancing the safety and navigability of the means of transportation in unpredictable indoor environments.