An Intelligent Embodied System for Enhancing Social Interaction
Embodying Intelligent Behavior in Social Context | Intelligent algorithm, Machine learning, AI
Team: Mitchell Ansems, Almar Stappen, Jing-cai liu, Solveig Skavnes, Florijne Merton


There are many situations in which the setting is ripe for individuals to approach and get to know others outside of their social circle. Sometimes this might even be desired, like in a festive setting, a social event or business occasions. However, many have a hard time when it comes to approaching new people, as well as being bound by the presence of a pre-existing social group. For these types of occasions, we look to provide an embodied intelligent algorithm that can match individuals that could potentially connect well, that otherwise may have never taken the time to talk to one another. It is hypothesized that an embodied intelligent algorithm may heighten the likelihood of initial contact between such parties. According to Turner (1988), we need to ‘trust’ others to be motivated to initiate interaction, where trust is higher when there is a feeling of security and when there is sufficient information about the other person [1]. The design here is supposed to play a role in kick-starting a conversation, as well as providing the certainty that both parties have some common ground they can bond over, although they may not know exactly what this common ground will be.



RecoFriend is an intelligent wristband that can be worn at large events in which people could meet new social contacts. It facilitates social interaction by breaking the ice in starting a conversation with a stranger. The goal here is to increase people’s chances of meeting others that might have similar interest, but that they would not regularly talk to. RecoFriend plays the role of a recommendation device that shows individuals with a similar profile to your own. The wristband displays a color that is linked to a specific class of people. One class of people may have the same interest, personality, lifestyle, or any other feature that people could bond over. Your profile or class is retrieved from survey data that you fill out at the beginning of the event. This process is arranged and guided by the party that acquired or rented the wristbands, which will usually be the organizer of said social event.
In terms of matching, the recommendations of the wristband will change after having conversations with another person meaning that the system learns over time what one’s preferences are and what they might find important. During each conversation, the wristband measures the length of the conversation as an indicator of match success. This is done by starting a count after a handshake and end the measuring when the distance between the persons is far enough, that it can be concluded that their interactions ended. This additional data from different conversations could then potentially change your class or profile to more accurately match you with future partners.


Every data point represents a person/bracelet. The implemented algorithm supports up to five different questions to be asked of a user. Each question represents a feature in the feature space, making the maximum feature space five-dimensional. Since five dimensions are not possible to illustrate, the graphic shows only the first two. The reason there are more data points than circles in the graphic is that several data points overlap in the first two dimensions.



The goal of this project was to find out whether or not the implementation of an embodied intelligent algorithm would be a suitable approach for supporting social communication initiation. Because of the scope of this project as well as being advised to do so by the case owner, testing with actual users was refrained from and conclusions drawn were limited to the theoretical capabilities of the prototype.

Conversation success

As mentioned before, granting two users the same classification is not a guarantee of conversation success. Where the same class might mean that they have similar interests in the algorithmic sense, this may not always translate to similar real-world opinions.

Measurement of the conversation

It might happen that one is put into a type of classification that does a bad job of representing their true character. When the conversation is long, that could mean that the two individuals have had a good conversation with each other. However, this is not always the case. For example, in some conversations, one person is speaking continually while the other person is only listening.

Physical limitations

Design shortcoming like these was chosen purposefully, considering the timeframe of the case and the privacy of its user group. Measurements on success could be improved by including a number of biosensing parameters, however, this heavily interferes with the casual interaction goals of the user group and invades on their privacy as well.


+ Math Data & Computing
+ Technology & Realization
Throughout this lecture, I have improved my skills in Math Data & Computing and Technology & Realization. Before this elective, I had no experience in designing within the subject AI. In this elective, I have learned different ways to design an intelligent system. I have improved my general understanding of the methodologies of intelligent algorithms, how machine learning works, and the differences between supervised and unsupervised learnings [2]. Although I think that I still need to learn a lot about this topic and to try-out more in order to design a nicely working design, I think the knowledge from this lecture could already help me to understand AI and related other works.


1. Turner, J. H. (1988). A theory of social interaction. Stanford University Press.
2. Love, B. C. (2002). Comparing supervised and unsupervised category learning. Psychonomic bulletin & review9(4), 829-835.
Theme: Overlay by Kaira