Toyota announced its robotics institute in Cambridge in 2015, as well as a second institute and headquarters in Palo Alto, California. In its home country of Japan – as in the United States and other wealthy countries – the population is aging rapidly. The company hopes to build robots that can help people continue to live independent lives as they age.

The Cambridge lab has dozens of robots working on tasks like peeling vegetables, using hand mixers, preparing snacks and flipping pancakes. Language models prove useful because they contain information about the physical world, helping robots make sense of the objects in front of them and how they can be used.

It’s important to note that despite many demos slick enough to impress a casual visitor, bots still make a lot of mistakes. Like earlier versions of the model behind ChatGPT, they can fluctuate between human-looking and strange errors. I saw one robot effortlessly operate a hand mixer and another struggle to grip a bottle cap.

Toyota isn’t the only big tech company hoping to use language models to advance robotics research. Last week, for example, a team at Google DeepMind recently unveiled Auto-R, software that uses a large language model to help robots determine what tasks they could realistically and safely perform in the real world.

Progress is also being made on the hardware needed to advance robot learning. Last week, a Stanford University group led by Chelsea Finn released videos of a low-cost mobile teleoperated robotic system called ALOHA. They argue that being mobile allows the robot to tackle a wider range of tasks, giving it a wider range of experiences from which to learn than a system locked in one location.

And while it’s easy to get dazzled by robot demonstration videos, the ALOHA team was kind enough to post a highlight reel failure modes showing the robot fumbling, breaking and knocking over objects. Hopefully another robot will learn to clean after this.