Training robots how to laugh is humorously hard
A laughing robot might be seen as menacing, but would it appear any less scary if it laughed at the right time during a conversation? That’s a theory that scientists have recently tested. A group of researchers from Kyoto University in Japan have recreated a laughing robot named Erica powered by an AI system that focuses on conversation.
Since laughter is a normal part of human dialogue, they argued, it might be useful to see how humans respond to talkative robots that they can also laugh with. Their findings were published in the journal last week Frontiers in robotics and AI.
Artificial intelligence is good at logic, but laughing? Not as much. First of all, the researchers recognized that there are various reasons why people laugh – and that complicates things even more. To make it easier for the AI system, they generalized laughter into two categories: shared social laughter, when the AI laughs in response to a human laugh, and solo happy laughter, when the robot laughs in response to a topic or while talking .
The researchers trained the AI model on how and when to laugh by allowing it to participate in a sort of speed-dating with male university students. Erica was remotely controlled by a female actress who spoke into a microphone and controlled physical movements like head nods and other gestures.
The chats lasted 10 to 15 minutes and the data was taken from 82 conversations. The researchers recorded the conversations with microphones and cameras and annotated them based on when people’s social laughter and solo laughter occurred and how that laughter differed. This data was then used to train the AI system to teach it when to use and what type of laughter. They then applied their shared laughter algorithm to existing conversation software and asked 130 volunteers to listen and rate how well the robot simulated empathy, understanding, and human likeness.
Overall, the researchers found that in situations where shared laughter was appropriate, the researchers found that Erica and her algorithm did a good job of convincing people to pay attention to what was being said. But it had some flaws and limitations. Erica was good at responding to laughter, but didn’t really have a sense of when to laugh on her own. The researchers wrote in their discussion that this may be because it’s easier to learn to respond to a prompt than it is to actually understand why the content of the conversation might be funny.
Whether or not Erica actually gets human humor is just part of a big project roboticists and engineers are involved in: giving robots social skills. Since 2017, scientists have been working on how to make a robot laugh convincingly (big tech companies like Microsoft, IBM and Meta are also interested in this). A month ago, Italian engineers unveiled a bartender robot capable of making small talk (unfortunately, due to privacy concerns, it’s shelved for the foreseeable future). The idea is that robots’ facial expressions, body language, language, and the ability to understand and respond to people’s demands make them more engaging and better at day-to-day interactions.
But ultimately, there can be a slippery slope from a social exchange that feels natural to an eerie valley scenario. There are also ethical concerns with robots being too credible. Still, there are practical reasons to keep working in this area: Making talking robots less scary and more approachable by giving them the right human-like qualities will be particularly useful, according to experts, to one day integrate them into healthcare, hospitality or other service-oriented industries.