As we’re learning more and more everyday, artificial intelligence can do a lot. AI can inform predictive text on smartphones, control home security systems, autonomous vehicles and even help operate prosthetic devices. But can AI do improv comedy?
A trio of researchers — two based in Edmonton at the University of Alberta — wanted to find out how a robot, running a custom language program and armed with substantial pop culture references, would fare on stage.
Blueberry debuted at the Fringe Festival this past summer.
Blueberry was built using a robot frame from a company called SoftBank, which runs a software called ALEX — or the artificial language experience — created by Kory Mathewson, Riley Dawson and Piotr Mirowski.
“We moved around a bunch to test the bounds and limitations of all the different joints,” Mathewson, a PhD candidate, explained. “I trained the system on a bunch of dialogue from movies because like improv actors, the robot can pull ideas, thoughts, quotes from movies and then piece them together in certain ways.”
The team then decided to test Blueberry’s AI in a very public setting: on stage, live, at the Fringe.
“It went really well,” Mathewson said. “The critical review was very positive. We sold out shows, we had a lot of great audience feedback throughout the shows.
During the shows, audience members were invited to interact with Blueberry and each performance had a different improv actor join the troupe on stage.
“We had guests every night of the Fringe and every show we’ve done we’ve brought in improvisers to work with the system and they loved doing it. It’s like a world of inspiration for them,” Mathewson said.
“It’s all because the system has a lot of randomness to it and it challenges the performers to do well.”
Audiences described the AI-human show as “compelling” while being “the most technically dangerous show at the Fringe.” So many elements are left to chance and there’s so much risk. But, for Mathewson, it was the perfect marriage.
“I also love improv so it was a way to merge these two things that I love and test one with the other.
“The themes of AI and improv are aligned. They’re both kind of built on this idea that you learn the most from failure and so if you’re able to fail and learn a lot from it, then you can succeed on the next chance you have to act or interact with the world.”
Taking their show on the road allowed the team to show family and friends a different — and fun — application of their work. They also hope it generates more interest in computer science and AI.
Watch below: AI-enabled robot Blueberry answers reporter Emily Mertz’s questions about performing at the Edmonton Fringe Festival. PhD candidate Kory Mathewson also weighs in on the process and response.
“If we find ways to translate the more scientific, nuanced things that we’re doing to a broader audience then that gets a lot of people, more young people, really excited about what we’re doing and kind of encourages them to keep pursuing stuff like this and not be too intimidated,” Dawson said.
The software engineer was on stage for all the Fringe performances, behind a computer desk that helped direct — or not — Blueberry’s actions.
“As people were taking their seats and stuff I would actually just start talking to the audience through the robot,” Dawson said. “At that time, they have no idea if I’m doing that or what my involvement is.
“Seeing the delight people had when they reacted to what the robot was doing, it was really satisfying.”
Mathewson said he was also surprised how interested the audience was in learning how the robot works; not just seeing it in action.
“I’m amazed how amazed people are about robotics still. It feels so second nature to us in these types of labs because we sit with and tool around… but to see the joy that these kinds of systems bring to people and that moment of awed-ness, that awed look they give when we’re traipsing down the valley of uncertainty and unfamiliarity is so fun, so exciting.”
The team will use the improvization experience and how Blueberry reacted as part of their studies. Feedback collected throughout the performances from actors and audience members will be compiled for their research and later published.
“A lot of the science that we’re doing here is focused on improving human-machine communication so hopefully some of the findings will contribute to building better interactions for humans to communicate with, control and tune the AI systems they interact with,” Mathewson said.
And, there’s a second act.
Another series of HumanMachine performances called Improbotics are coming to the Citadel Theatre in January. Click here for more information and to buy tickets.