Researchers at Google said that by using large language models, it can make robots respond better to complex and abstract requests from humans.
Google is combining large language models with its Everyday Robots so they can help humans with a broader range of everyday tasks.
Everyday Robots are being developed by X, the experimental division of Google’s parent company Alphabet. They were introduced to Google offices last year to help clean up around the workplace.
Researchers at Google said that currently most robotic systems can only respond to short, specific commands, such as “pick up an apple”. The team believes large language models hold the key to improve these machines.
Large language models are natural language processing systems that are trained on a massive volume of text.
By using large language models such Google Research’s PaLM, it is possible to have the robot interpret the right response to more general commands or react to a situation properly.
For example, if someone says “I spilled my drink, can you help?”, the robot can make the decision to get a sponge and clean the mess, rather than needing to be told “pick up a sponge”.
“We thought that combining this advanced language model with a robot that can learn by itself, will allow us to leverage the benefits of each,” said Google research scientist Karol Hausman in a video.
The team behind the system, dubbed PaLM-SayCan, seem happy with the progress so far. From testing 101 tasks in an office kitchen environment, Google said the robot chose the right skills to perform from a request 84pc of the time, while executing it successfully 74pc of the time.
“We are still far from this becoming a household staple and there is a long road ahead before we can communicate with robots more naturally,” Hausman said.
“However, we are excited to see the possibilities as we continue to explore new approaches to train more human-centred robots.”
10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.