Vehicle-Driving in Spatially Configured Workspace by a Full-Sized Humanoid Platform, HART

Abstract This paper presents the development of an autonomous vehicle-driving control system for the full-sized humanoid platform, HART (Human Assistive RoboT), a kinematically resizable biped which can drive various types of vehicles by changing its mechanical dimension upon given circumstances. First, the landmark-based wayfinding navigation control architecture, which enables HART to drive the assigned vehicle in spatially configured environments, is described. In the architecture, the guideline trajectory, which consists of waypoints that link start and goal locations, was initially generated from the path-planning process and was then fed into the landmark recognition module. The trajectory information and GNSS data were used as main references of the lane and intersection detection process in the recognition module along with real-time incoming image data. The module then generated the recommended motions for HART to manipulate main control inputs of the vehicle, adaptively to the recognized intersections, which are landmark waypoints of the path. Next, the CNN (convolutional neural network) based intersection classifier, which can strengthen the perception capability of the landmark recognition module, is introduced with investigations on various hyper-parameters and structures of the network, and also a performance comparison with other models. Lastly, the autonomous driving capability of HART, with the developed navigation system, was tested and evaluated through experimentation in a real-world setting.