Plan Explainability and Predictability for Cobots

Robots are becoming pervasive in human populated environments. A desirable capability of these robots (cobots) is to respond to goal-oriented commands by autonomously constructing plans. However, such autonomy can add significant cognitive load and even potentially introduce safety risks to the humans when robots choose their plans unexpectedly. As a result, for cobots to be more helpful, one important requirement is for them to synthesize plans that do not {\it surprise} the humans. While there are previous works that studied socially acceptable robots which discuss ``natural ways'' for cobots to interact with humans, there still lacks a general solution, especially for cobots that can construct their own plans. In this paper, we introduce the notions of plan {\it explainability} and {\it predictability}. To compute these measures, first, we postulate that humans understand robot plans by associating high level tasks with robot actions, which can be considered as a labeling process. We learn the labeling scheme of humans for robot plans from training examples using conditional random fields (CRFs). Then, we use the learned model to label a new plan to compute its explainability and predictability. These measures can be used by cobots to proactively choose plans, or directly incorporated into the planning process to generate plans that are more explainable and predictable. We provide an evaluation on a synthetic dataset to demonstrate the effectiveness of our approach.