强化学习
计算机科学
杠杆(统计)
机器人
规划师
人工智能
运动规划
控制器(灌溉)
路径(计算)
参数统计
机器学习
农学
数学
生物
统计
程序设计语言
作者
Nathan Hemming,Vineetha Menon
标识
DOI:10.1109/ictai59109.2023.00049
摘要
Robust and efficient navigation methods are needed to allow a robot to work independent of human control. Many methods leverage models of the world and robot, maps of the world, and graphs to generate navigation plans. Each of these tasks takes additional time away form the limited computational budget of real time systems present on robots. The methodology presented is called declarative re-planning (DRP) and is graph free, model free, and map free. The proposed method is able to achieve orders of magnitude increase in learning efficiency and to navigate and reach its assigned destination location 100% of the time in an experiment set. DRP leverages an hierarchical interface between a traditional path planner, a deep reinforcement learning based policy, and a classical model controller to achieve its performance. The top level create waypoints that serve as sub goals. The middle level generates a parametric path to reach the subgoal. The lowest level uses a model controller to follow the path. Each operates on a different time scale and is concerned with different aspects of the navigation task. All three levels of the hierarchy work together to obtain robust navigation capabilities.
科研通智能强力驱动
Strongly Powered by AbleSci AI