RoboBoom
CITRIS and the Banatao Institute· December 3, 2021· 52:46

Multi-Party Human-Robot Interaction

Marynel Vázquez, a Yale professor who leads a multiparty human-robot interaction lab, discusses how groups respond to robots and how structured machine learning can help robots understand and join those groups.

In this talk

  1. 3:00

    Bystanders respond to robot mistreatment

    Vázquez describes experiments testing whether a robot could prompt a bystander to intervene when a confederate mistreated it. In one study, 27 of 48 participants made disparaging remarks, but only two directly challenged the confederate. In a follow-up, other robots turned toward the abused robot and showed sadness; participants intervened more often than when the robots stayed still.

  2. 12:26

    Group behavior changes the interaction

    Vázquez argues that group interactions bring effects that do not appear in one-on-one encounters. Robot gaze can affect participation, and robots may influence bonds between people. The experiments also leave open why bystanders intervened more: empathy, conformity, or another social mechanism.

  3. 15:08

    Models for detecting conversational groups

    Vázquez shifts from scripted studies to robot autonomy, advocating structured data-driven methods. Her earlier work used F-formations, such as face-to-face or circular arrangements, to estimate who was conversing with whom, but those rules struggled with context such as queues, walking, and elevators. Deep Sets and graph neural networks can handle variable group sizes and unordered people; graph models improved group detection as the available node features grew.

  4. 37:12

    Predicting where a robot can join

    Vázquez compares a geometric method for placing a robot within a conversational group with a graph-based generative model. In a study, the GAN was rated more favorably for groups of four to six, while the geometric method did better for groups of two or three. The gap exposed a flaw in simulated training data: it used the same group radius regardless of group size.

  5. 43:59

    Scaling data collection for social navigation

    Vázquez discusses using existing human interaction datasets, long-term robot deployments, and browser-based simulations to gather data. Her lab connects Unity simulations to ROS, runs them in the cloud, and lets participants control an avatar through a browser survey. Interactive simulations produced different feedback from video surveys, and the team had not yet established which better matched real-world responses.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks