Very little work has been carried out using emotions and their facial expressions in the context of human-robot interaction. Cynthia Breazeal has developed Kismet as a testbed for learning social interactions in situations involving an infant (the robot) and her caretaker (a human). Kismet is a head with active stereo vision and configurable facial features--eyelids, ears, eyebrows, and a mouth. Humans can interact with it either by direct face-to-face exchange or by showing it a toy. In [1], Breazeal reports on some experiments where Kismet uses nine different facial expressions to manipulate its human caretaker into satisfying its internal drives--a social drive, a stimulation drive, and a fatigue drive. Since the focus of this work is on social learning, no experiments to test the expressive aspects of the interaction have been reported so far. However, given that we have taken almost opposite working hypotheses concerning expressive facial features (a rather sophisticated, configurable set of features in her case, versus a minimalist one in ours), it would be highly interesting to perform common tests on human recognition of the robots' expressions and to compare our results.
Minerva, developed by Sebastian Thrun, is a mobile robot that gives guided tours to visitors of the Smithsonian's Museum of American History [13]. It displays emotional states--neutral, happy, sad, and angry--using a caricaturized face and simple speech. Emotional states arise as a consequence of travel-related interaction (e.g., anger results from its inability to move due to the presence of people), and their expressions aim at affecting this interaction towards achieving the robot's goals--traveling from one exhibit to the next one, engaging people's attention when describing an exhibit, and attracting people to participate in a new tour. Although [13] reports very successful interactions that the authors attribute to empathetic feelings in people, it also states that emotions in Minerva are purely a means to an end and not an integral part of the robot's architecture and interface.
A surprising experiment, although using a computer instead of a robot, was conducted by Clark Elliott [7] to test the computer's ability to express emotions by having humans recognize them. The computer used both caricaturized facial expressions and voice inflection to convey different emotional states while saying sentences devoided of emotional content. As a control, he had an actor say the same sentences and express the same emotions. It turned out that humans performed substantially better when recognizing the emotions expressed by the computer (70% of success) than those expressed by the actor (50% of success). Elliott suggests that these results might be partly due to the use of caricaturized expressions.