Pepper as a Social Assistant

The goal of this project was to develop a flexible, modular, and user-friendly software for researchers working with the Pepper robot. Build-ing on prior work by our clients, we developed a software application that provides researchers with an intuitive user interface, allowing them to shape Pepper’s behavior using a no-code visual programming envi-ronment. First, a researcher builds a program by creating a sequence of phases, each with their own goals, norms, and trigger-driven actions. Then, once the researcher chooses to run their program, the interface switches to monitoring mode. Here, researchers can obtain any necessary information about their experiment and influence Pepper’s ongoing con-versation. At the same time, our software converts their program into AgentSpeak, which is then used by our Agentic belief-desire-intention-based backend to drive Pepper’s interactions, ensuring they feel as natu-ral and organic as possible. To achieve this, the backend utilizes a locally running large language model to generate responses. These responses are shaped by the influence of a multitude of agents that collectively provide information about the ongoing interaction. This approach enables Pep-per, for example, to extract beliefs from the dialogue and automatically transition between phases of the interaction by determining whether the goals for the current phase have been achieved. The result is a tool that enables researchers to precisely define natural and organic interactions between man and machine.