Learning complex task structures from demonstrations
Our goal is to learn the structure of complex multi-step tasks from demonstration of human experts.
The goal of this project is to develop models and algorithms for inferring the intent behind a human expert’s actions and his/her schema for accomplishing complex tasks, e.g. military tactical missions and emergency response deployments. Traditional approaches to learning from demonstration are not well suited to learning multiple schemas and their transition and constraint relationships. For example, a tactical air mission has multiple phases but these phases are not typically flown in a fixed order. Our goal is to learn the component structures and their transition models from hybrid continuous and discrete data streams in both supervised and unsupervised learning settings. The learned schemas can be used reconstruct the mission story, and can be used in an intelligent system that performs critical review of the missions to improve future performance.