Learning contiguity-based hierarchical task models from demonstration
Rossignac-Milon, Leo Thomas
MetadataShow full item record
We propose an incremental approach for learning a hierarchical task model from a series of demonstrations, where each demonstration is a permutation of a fixed number of different actions. Our hierarchical Task Execution Model, called TEM, is a tree, where each leaf represents an action and each node represents a composite action (or subtask). We distinguish three types of composite nodes (s-group: sequential, r-group: reversible, and u-group: unordered). Although the sub-task children of a node must always be executed as a contiguous (uninterrupted) sequence, the valid orders for that sequence depend on the node type. Hence, a TEM captures a well-defined set of contiguity and ordering constraints. TEM may be used to test quickly whether a candidate plan of actions is compatible with the task model and also to provide a list of valid actions at any step during the lazy execution of a task. We propose an incremental algorithm that takes as input the current TEM learned from previous demonstrations and a new demonstration in order to produce a new TEM.