Autonomous ground robot motion decision based on supervised learning

Abstract An embodied robot is necessary to extend localization and mapping capabilities to action decision-making for autonomous tasks. Such tasks usually require ground robots to explore in unknown environments. This paper addresses motion planning with a compound visuomotor policy network, which allows the robot to actively localize itself and build a global map. We emphasize that localization derived by Simultaneous Localization and Mapping (SLAM) could also feedback to visual-motion decisions by offering spatial knowledge that could be learned by supervised learning. This spatial knowledge is proposed to be exploited by a novel position information fusion strategy for visual observations. This paper features the mutually beneficial mechanism between the tightly coupled multi-sensor fusion SLAM module and the visual-motion Planning and Control Module (PNCM), which is composed of a visual encoding Transformer and a diffusion network. The effectiveness and efficiency of the proposed idea are validated by simulation experiments.