HiBRIDGE: A Hierarchical Bayesian Neural Network Framework for Interpretable Dialogue Management in Group-Robot Interaction

Imagine a robot walking into a room where three people are already talking. To join the conversation without causing confusion or awkwardness, the machine must make a split-second decision: who should it speak to, and what should it say? Should it ask a question to the person on the left to keep a topic going? Should it offer a joke to the whole group? Or should it simply listen? In the complex world of human-robot interaction, there is rarely just one "correct" answer. Multiple choices can seem reasonable at the same time, and the right move depends entirely on the shifting context of the moment. For a robot to be a true partner in these group settings, it needs more than just a script; it needs a way to weigh these possibilities, learn from limited experience, and explain its choices in a way that humans can understand.

This is the challenge tackled by a team of researchers who developed a new system called HiBRIDGE. Their work focuses on the "brain" of a social robot—the part that decides how to behave in a group. Instead of relying on a single, massive calculation that tries to guess the perfect response all at once, the researchers broke the decision down into smaller, logical steps. They built a system that first decides if the robot should speak at all, then decides whether to address the whole group or just one person, and finally decides what kind of thing to say, such as asking a question or making a statement. Crucially, they designed this system to handle uncertainty. Because real-world interactions are messy and data is hard to collect, the system uses a method that acknowledges it might not be 100% sure of the answer, allowing it to learn effectively even from a small number of examples.

The researchers tested this new framework using three different sets of recorded conversations. In one scenario, a robot entertained shoppers in a mall with quizzes and jokes; in another, it answered questions in a hospital memory clinic; and in a third, it helped moderate a negotiation game between roommates. When they compared their new system against existing methods, including those powered by large language models, HiBRIDGE performed better. It was particularly effective at predicting the right behavior when the task was difficult and the data was scarce. The study found that the approach which treats the robot's decision-making as a series of probabilistic steps—essentially weighing the odds of different outcomes—consistently outperformed systems that tried to make a single, fixed prediction. This suggests that admitting uncertainty can actually make a robot smarter in unpredictable social situations.

Beyond just being more accurate, the team wanted to know if this step-by-step structure made the robot easier to understand. To test this, they ran an online study with 20 participants. They showed people video clips of robot interactions and provided explanations for why the robot acted the way it did. Some explanations came from the new, structured system, while others came from a simpler, "flat" system that made all its decisions at once. The results showed a clear preference: people found the explanations from the structured system more helpful and easier to follow. When asked to choose between the two, participants consistently picked the explanations that revealed the robot's intermediate thought process, such as "the robot decided to ask a question to the whole group because the conversation had stalled." This indicates that showing the "why" behind a decision, rather than just the final result, helps humans trust and understand the machine.

To see if this worked in the real world, the researchers built a fully autonomous robot and tested it with 12 people in a face-to-face setting. The robot, equipped with cameras and microphones, listened to a group conversation, decided when to speak, and then generated its own words and gestures. The study found that the system worked smoothly in real time. Participants rated the robot's behavior as appropriate and useful, regardless of whether the robot was using the complex, structured decision-making model or a simpler version. While the structured model did not statistically outperform the simpler one in every single rating, it was consistently perceived as slightly better at managing the flow of conversation and including everyone. This real-world test proved that the new framework is not just a theoretical exercise but a practical tool that can run on a robot, making split-second decisions while maintaining a natural and engaging presence.

The work highlights a dual benefit in how we design intelligent machines. The mathematical approach that handles uncertainty helps the robot learn faster and perform better when data is limited, which is common in real-world robotics. Meanwhile, the step-by-step structure of the decision-making process provides a clear window into the robot's mind, making its actions more transparent to the people it interacts with. By combining these two elements, the researchers have created a foundation for robots that are not only capable of navigating complex social groups but are also able to explain their choices in a way that feels logical and human. This moves the field closer to a future where robots can participate in our conversations not just as tools, but as understandable partners.