Anyone who has ever shared a laugh with a friend knows how deep a connection humor can evoke. So it makes sense that future companion robots are more likely to win our trust and our affection if they can laugh with us. They still need to know how to laugh properly depending on the situation.
That is why Japanese researchers are trying to teach robots to laugh at the right time and in the right manner. It turns out that teaching an AI to laugh is not as easy as teaching it to answer a phone call. “Systems that try to mimic everyday conversations still don’t know when to laugh,” according to a study published in the journal Frontiers in Robotics and AI.
Don’t laugh for a laugh
The team is working on an AI dialogue system based on collaborative laughter to make human-robot communication more natural. The researchers plan to integrate this system into existing conversational software for robots that are already learning to detect emotions and deal with complex open-ended questions.
“We believe that one of the important functions of conversational AI is empathy,” explains Koji Inoue, associate professor of computer science at Kyoto University (Japan) and co-author of the study. “The conversation, of course, is multimodal, and not only about the correct answer. Therefore, we decided that one of the ways the robot can empathize with users is to share their laughter. »
The bottom line is that the system not only recognizes laughter, but also decides whether to laugh back, and then chooses the type of laughter that best suits the circumstances.
To collect training data on the frequency and types of shared laughter, the team used Erika, a humanoid robot developed by Japanese scientists Hiroshi Ishiguro and Kohei Ogawa, as a platform to study human-robot interaction. Erika understands natural language, has a synthesized human voice, and can blink and move her eyes when listening to people speak.
The researchers recorded conversations between students at Kyoto University who took turns chatting with Erica while amateur actresses in another room controlled the robot through a microphone. The scientists chose this configuration knowing that there would naturally be differences between how people talk to each other and how they talk to robots, even those controlled by another person. “We wanted, as far as possible, to train the laughter model in conditions that are close to those of a real human-robot interaction,” says Divesh Lala, a researcher at Kyoto University and another co-author of the study.
10 to 20 years to natural conversation with a robot
Based on these interactions, the researchers created four audio dialogues between humans and Erica, who was programmed to respond to conversations with varying degrees of laughter, from deadpan to frequent giggles. The volunteers then rated these sideshows for empathy, naturalness, human likeness, and understanding.
The collaborative laugh scenarios performed better than those in which Erica never laughs or laughs every time she detects human laughter, without using the other two subsystems for context and response filtering.
However, while robots are becoming more and more realistic, in sometimes frightening ways, roboticists recognize that imbuing them with human traits poses challenges beyond programming. “It will probably be 10 to 20 years before we can finally talk to a robot in an informal setting, like a friend,” admitted Koji Inoue.
CNET.com article adapted by CNETFrance