Leggi in italiano

iCub, the small humanoid created by the Italian Institute of Technology in Genoa. Credit: Michele D'Ottavio / Alamy Stock Photo.

Variability is one of the fundamental features of human behavior. Researchers believe that it could be the result of evolution because it makes us more unpredictable and less vulnerable to predators.

This same feature could also play a role in our ability to tell what is human, and what is not. Several experiments have shown that when humanoid robots exhibit human-like variability in response times or motion patterns, we perceive them as more human-like.

In a study published in Science Robotics1, researchers have observed this same effect when the human and the robot are performing a shared activity. “To evaluate the impact of behavioural variability in the attribution of humanness to a robot, in our experiment the robot was either teleoperated by another human or controlled by a computer”, says Agnieszka Wykowska, senior researcher at the Italian Institute of Technology (IIT) in Genoa, and the coordinator of the study.

The research has also shown that the effect applies even when the variability of the robot’s behaviour does not closely resemble the human one, if it falls in the same range. “Depending on the context and on the function that the robot needs to perform, roboticists can endow their machines with a different degree of humanness by modulating the variability of their behaviors,” Wykowska adds.

For the experiment, a human participant was seated in a booth in front of a screen, next to an iCub humanoid robot. The set-up was replicated in an adjacent booth. In one booth, when a red square appeared on the screen, the participant had to press the button in front of them. When a green square appeared, it was down to the iCub had to press its key. The iCub’s reaction times were controlled either by the human in the other booth, where the task was performed in the opposite way, or by a pre-programmed computer algorithm. This algorithm reproduced the same range of variability of human reaction times measured by the authors in a previous study, but without the same shape of the statistical distribution2.

The participants did not know the control configuration of their iCub, nor when they were teleoperating the robot in the adjacent booth. “This ensures that we are studying a real-time interaction”, explains Francesca Ciardo, another researcher at IIT and first author of the study.

The same control configuration, computerized or teleoperated, was kept for sequences of 100 trials. At the end of each sequence, the participants were asked if they believed the iCub was controlled by either a human or a computer.

Most of the times, the participants could correctly guess when the iCub was indeed teleoperated by another human, signaling that we are sensible not only to the range of the reaction times’ variability, but also to the shape of its distribution. Instead, at the end of sequences where the iCub was controlled by the computer, participants gave the right answer only around 50 per cent of the times, the same probability one would expect if they were answering randomly.

The researchers also measured the coordination between the human and the robot, by looking at the degree of correlation between their reaction times, and found that it was higher when the iCub was controlled by the computer. “This could be due to the fact that the robot’s reaction times were more predictable”, Wykowska says. It could be desirable to adopt this kind of variability when robots are used more as tools than as real collaborators, for example in robot-assisted surgeries. “On the contrary, we might sacrifice efficiency in performance when acting in social contexts”, Wykowska adds.