An AI enabling robots to do chores like making the bed or cleaning up spills in homes it has never seen before could allow many more robots to become generally useful, its creators say.
Large language models (LLMs) that power tools like ChatGPT have improved robots’ ability to carry out spoken requests. However, most robots work well only in environments in which they have been trained; their performance quality sharply falls when confronted with new and unfamiliar spaces.
Now, at robotics company Physical Intelligence in San Francisco and his colleagues have developed an artificial intelligence model called π0.5 that allows robots to work in real homes that they have never seen before.
Advertisement
“The thing that we’re quite excited about is that we’re finally getting to the point where these models can generalise well enough that we can put robots, equipped with these kinds of systems, into real-world environments,” says Levine.
The model is trained using data from an unusually wide range of sources, including many different kinds of robots working in lab and home settings, as well as extensive data scraped from the web, such as image and object databases.
When Levine and his team tested the AI model on robots they had built using off-the-shelf components, 97 per cent of the training data came from sources other than the robot itself. “Because π0.5 can leverage other data sources, from the web and other kinds of robots, then it can have this broader generalisation,” says Levine.
The team tested how well the robots performed when instructed to do chores such as putting plates in the sink, putting shopping away in drawers and placing dirty clothes in a basket.
They didn’t work perfectly every time. “When it comes to the practical utility of this model, it’s definitely not there yet, in the sense that this is not a home robot that somebody could buy and put in their home,” says Levine.
But he says that the robots’ performance appears to improve in a predictable way as they are tested in increasing numbers of homes. This trend, called a scaling law, is similar to how language AI performance was found to predictably improve as more data was added.
For instance, when a robot gets to around 100 homes, it can perform a task as well as a robot that has been trained and tested in a specially designed testing facility. “What we have established, and that I think is very, very promising, is the beginnings of these kinds of scaling laws [for robots],” says Levine.
“The videos they show are quite impressive,” says at the University of Lisbon in Portugal. However, the robots perform their tasks at least five times slower than a human, sometimes taking 10 to 15 minutes to finish, which could be difficult to improve, he says.
“Doing [tasks] fast is a real problem for robots,” says Lima. “It has been, and it will be still for quite some time, because the robots need time to do the computation, especially if they do it on board.”
arXiv,
