In this project, we produced a framework for autonomy for the Tiger Taxi, formerly known as the Autonomous People Mover. Whereas previous teams have focused on local cost-based or ad-hoc navigation, we have attended to the lack of global mapping and localization. Through a set of bag file-based and stage-based simulation environments, as well as preliminary path planning, we have produced a lucrative platform for point-to-point autonomous navigation. A baseline navigation configuration for simple goals demonstrates the path planning potential of the framework. We utilize the probablistic Octomap mapping package to accumulate filtered laser scans. Global localization estimates supply the module with the transform required to align the point clouds with previously collected data. The ENet convolutional neural network is trained to classify both unsafe and safe regions of the observable terrain. Where the LiDAR fails to detect, e.g., subtle curbs, Safezone is able to compensate. The GUI facilitates goal-setting and debugging of the Tiger Taxi subsystems. A global reference frame indicates the cart's position with various windows for configuration, monitoring, mode-setting, and horn noises.
View documentation of the TigerTaxi project.