基于子任务的机器人潜在动作预测

Robot′s affordance prediction based on the subtask

以动态环境下的机器人导航为例,研究了机器人在任务复杂、物体随机出现等情形下的潜在动作预测方案.采用层次结构描述机器人的任务,提出了一种新的形式化描述模型,将潜在动作的影响范围从原子动作提升到子任务层次.提出的潜在动作预测框架集成了分层强化学习、状态抽象机制、任务图和物体属性.迷宫环境下的导航实验表明:机器人能够根据当前子任务、自身的感知能力和行为能力以及物体的动作属性来预测潜在动作,基于潜在动作的方案比传统方案的效率更高.

Affordance prediction in dynamic environment was investigated,where the task was complex and objects appeared randomly.A robot′s task was decomposed into hierarchical structure,and a novel formalization was proposed to describe the affordance at the subtask level.An affordance prediction framework integrating hierarchical reinforcement learning,state abstraction,task graph,and objects′attributes was presented.In the experiment,the humanoid robot NAO should navigate in a maze environment where different kinds of objects are placed arbitrarily,and finish each subtask in an optimal manner.The results indicate that the robot could predict the affordances based on the current subtask,its own perceptual and motor capabilities,and the objects′functional attributes.Compared with the traditional non-affordance approaches,the affordance-based method performs better and finishes the task in fewer steps.