Hierarchical Task Planning from Object Goal State for Human-Assist Robot

Abstract

Task planning is a very important ability for human-assist robots that are used for helping humans with their housework. Given a complex and unseen task, these robots must plan and take appropriate actions according to the situation at that time in the human living environment. In this work, we focus on a method for task planning to learn a policy to hierarchically decompose a given task from the object goal state and the observation at that time. In the learning-based design of a policy for task planning, the optimal policy can be acquired from demonstrations without requiring to explicitly define the precondition. And we think that Hierarchical policy, in which a given task is decomposed into sub-tasks, makes it easier to learn the policy because the number of choices of actions becomes smaller by decomposing a given task into sub-tasks. Existing methods to hierarchically decompose a given task require the demonstration of the task to be executed, but our method does not require a demonstration for unseen task. In order to enable to hierarchically decompose the task from the object goal state, we split the model into a module to decompose the task and a module to predict the argument which is necessary for the action, and use the hierarchical structure of the task in predicting the argument. Having learned the policy, we carry out the execution of the task. We demonstrate the effectiveness of our method on the BlockStacking and Table-Arrange tasks.