A Transformer-Based Approach for Choosing Actions in Social Robotics

The growing deployment of social robots requires the ability to adapt to the dynamic changes occurring in the real environments. These reactive behaviors, however, are often incapable of reasoning and predicting the effects of their actions in the next future. Therefore, they must be accompanied by forms of deliberative semantic/causal reasoning. The combination of the reactive and deliberative forms of reasoning, which resembles the dual process theory, raises the problem of entrusting tasks to the corresponding modules. Just as happens in biological systems, the tendency to assign activities, as much as possible, towards the lower abstraction layers, equips the systems with more responsive capabilities at the cost of making the reactive layers more difficult to implement. In this document, we will introduce an architecture that, inspired by the classic three-tier architecture, combines slow and fast forms of reasoning, allowing social robots to achieve complex and dynamic behaviors. Since entrusting tasks to the more reactive components complicates their implementation (e.g., it requires the definition of formal rules which may not adequately generalize to unforeseen scenarios), we aim to reduce the technicalities and, consequently, to facilitate to the developers the implementation of the reactive behaviors. By relying on recent achievements in natural language translation, we will describe our recent efforts to adopt Transformer-based architectures, allowing the replacement of formal rules with easier to write "stories", defined through sequences of perceived events and actions, entrusting the system with the task of learning behaviors by generalizing from them.