Researchers Mizuho Okitani and Yoshinobu Hagiwara have built a model that lets a service robot learn who its users are from a few observations, without anyone labelling them, and still recognise them when their appearance, background noise or what they say changes, according to a paper on arXiv. They demonstrated it on a Toyota Human Support Robot.
The model combines what the robot sees, hears and is told, using features from the CLIP and CLAP AI models, in a Bayesian model that also works out how many different people it has met.
In tests with six participants under eight conditions, including changed appearance, added noise and transcription errors, it identified people with an average accuracy of 0.89, about the same as a standard clustering method that was given the true number of people, which reached 0.88.
Background
Mizuho Okitani
Robotics researcher
Researcher at the Faculty of Science and Engineering of Soka University in Tokyo.



