Improving accountability and explainability in robots through encryption, large language models, and visual language models

Abstract

Mobile robots are becoming increasingly common in spaces shared with humans, raising concerns about safety and security. However, the lack of capabilities to justify the actions behind a particular behavior to non-expert users poses a significant challenge in building trustworthy interactions. Explaining why a robot performs an unexpected action is crucial to understanding the cause of a failure. Furthermore, these explanations need to be clear and accessible to non-technical users. This work depicts a Proof of Concept of an accountability and explainability engine for Robot Operating System (ROS)-based mobile robots. Our solution comprises two components. The first consists of a black box-like module to provide accountability, capturing the actions performed by the robot while guaranteeing faithful replay, data privacy, and authenticity. The second component generates natural language explanations by using data within the black box. The explainability component is based on the use of Large Language Models (LLMs) and Vision Language Models (VLMs), enabling the system to analyse textual logs with the added context of visual input. Initial results show that it is possible to obtain understandable, accurate, and precise explanations by using accountable data as context. This fact enhances the deployment of responsible, transparent, interpretable, and trustworthy agents, easing interaction between humans and robots.