Robots Are in Their GPT-2 Era

It’s no secret that AI-powered robots aren’t so good yet. They struggle with a broad range of simple tasks, from untangling cables to chopping vegetables. Nonetheless, morale is high among roboticists who are flush with venture cash as they work toward a “ChatGPT moment” when robots become broadly useful and available. 

Until then roboticists are happy to celebrate little milestones. I spent two days this week at Actuate, a robotics conference in San Francisco, along with 1,200 founders, developers and engineers, where I saw demonstrations and video footage showing signs of progress.

When Chelsea Finn, a Stanford University professor of computer science and electrical engineering, played a clip of a mechanical robot arm successfully making a latte, the audience broke into applause—no matter that it moved slowly and a human had to steam the milk. (Some robot baristas are already serving coffee, for example at the SFO airport, but software code controls their actions, rather than a single AI model like PI’s that is trained to perform a variety of tasks.)

Finn, who is also a co-founder of robotics software firm Physical Intelligence, said her field resembles language AI in the era following OpenAI’s release of GPT-2 in 2019, when large language models were beginning to perform a variety of tasks without being specially trained or fine-tuned to specific ones. PI’s latest model similarly works out of the box on some limited tasks, such as using a pair of robot arms with pinchers to fold laundry and cardboard boxes.