Winter Robotics Colloquium: Marynel Vázquez (Yale University)

In this talk

1. 2:42

   Social context shapes robot performance

   Vázquez argues that general-purpose robots need to handle relationships among people, robots, and their surroundings, not just physical tasks. She defines social context as the attributes of relevant agents, environments, and their connections that affect an interaction. A cleaner held by a person, for example, can change what they mean by asking a robot to clean the fridge.

2. 9:55

   Robots can influence group behavior

   Vázquez describes experiments on how people respond when a robot is mistreated. In a study with 48 participants, only two intervened when a single robot showed distress or shut down. In a later collaborative task, other robots expressing sadness toward the mistreated robot led more participants to intervene.

3. 22:46

   Feedback and success depend on context

   Vázquez says there is no single best measure of social robot navigation: priorities vary across settings such as warehouses and homes. Her group studied implicit feedback, including people’s movement and facial expressions, to predict how they perceived a robot. In one navigation study, a random forest performed around human levels on binary perception predictions, while VR data also helped predict responses in real-world trials.

4. 35:51

   Large models expand HRI methods

   Vázquez describes using large language models to interpret people’s reactions to robot navigation and infer goals from language and actions. In the BALI approach, a robot combines what a person says with observed actions, weighs possible goals, and can ask questions when uncertainty is high. She also expects more autonomous deployments and argues that graph-based models may help robot behavior generalize to groups larger than those seen in training.