Discovering task constraints through observation and active learning

摘要

Effective robot collaborators that work with humans require an understanding of the underlying constraint network of any joint task to be performed. Discovering this network allows an agent to more effectively plan around co-worker actions or unexpected changes in its environment. To maximize the practicality of collaborative robots in real-world scenarios, humans should not be assumed to have an abundance of either time, patience, or prior insight into the underlying structure of a task when relied upon to provide the training required to impart proficiency and understanding. This work introduces and experimentally validates two demonstration-based active learning strategies that a robot can utilize to accelerate context-free task comprehension. These strategies are derived from the action-space graph, a dual representation of a Semi-Markov Decision Process graph that acts as a constraint network and informs query generation.We present a pilot study showcasing the effectiveness of these active learning algorithms across three representative classes of task structure. Our results show an increased effectiveness of active learning when utilizing feature-based query strategies, especially in multi-instructor scenarios, achieving better task comprehension from a relatively small quantity of training demonstrations. We further validate our results by creating virtual instructors from a model of our pilot study participants, and applying it to a set of 12 more complex, real world food preparation tasks with similar results.