Multi-Abstractive Neural Controller: An Efficient Hierarchical Control Architecture for Interactive Driving

As learning-based methods make their way from perception systems to planning/control stacks, robot control have started enjoy the benefits that data-driven provide. Because directly affect motion of robot, methods, especially black box approaches, need be used with caution considering aspects such as stability and interpretability. In this letter, we describe a differentiable hierarchical architecture. The proposed representation, called multi-abstractive neural controller , uses input image transitions within novel discrete behavior planner (referred visual automaton generative network, or xmlns:xlink="http://www.w3.org/1999/xlink">vAGN ). output vAGN controls parameters set dynamic movement primitives which provides system controls. We train controller real-world driving data via cloning show improved explainability, sample efficiency, similarity human driving.