Explainable reasoning for remote autonomous agents

Abstract

Remote autonomous robots are increasingly deployed for demanding tasks such as underwater exploration and pipeline inspection, providing valuable ecological insights and generating commercial benefits. However, human-in-the-loop applications in this domain face significant challenges, including a lack of direct supervision, bandwidth limitations, and limited technical understanding of the underlying autonomous systems. Ensuring situational awareness and trust is critical for the broader adoption of these technologies. This research project addresses these challenges by developing novel methodologies for transparent and explainable autonomy. The work focuses on two primary objectives: generating explanation content and effectively communicating it through natural language. To achieve the first objective, domain knowledge and robot state fusion are employed, alongside the creation of simplified autonomy models using surrogate techniques. For the second objective, the communication of explanation content is explored using both template-based and language model-based approaches, supporting causal, counterfactual, and contrastive explanations. User preferences for these explanation styles are evaluated, and the effectiveness of model-based explanations is compared to that of template-based alternatives. The findings demonstrate satisfactory performance in approximating autonomy using both surrogate and language models. Moreover, this work identifies the explanation styles that most significantly enhance situational awareness. These results contribute to the advancement of transparent and explainable autonomy, facilitating greater trust and adoption of remote autonomous robots in challenging applications.