Last week, I visited a robot demonstration just 15 minutes from my home, where an AI-powered robotic arm performed an impressive range of tasks.
I visited the offices of startup Generalist AI in Cambridge, Massachusetts. There, engineers showed me a robotic arm stacking cups and placing blocks into a bowl. I was surprised by how quickly the robot understood each task. Its behavior was remarkably similar to the flexible problem-solving abilities of a human.
The robot learned several tasks after watching a short instructional video, and it had not received specialized training for any of them. One of the most impressive demonstrations involved sweeping blocks into a bowl with a dustpan and brush. After the brush was removed, the robot improvised by using the dustpan itself to push the blocks into the bowl.
Courtesy of Generalist AI
In another demonstration, a two-armed robot watched a video of someone unzipping a wallet and removing banknotes. The robot then unzipped a different type of wallet and carefully took out the banknotes. When it struggled to grasp one of them, it switched from its right gripper to its left gripper to find a better angle. “Yes,” one of the engineers told me. “I’ve never seen anything like that before.”
“This is exactly the kind of thing that people were really excited about with GPT-3,” Generalist AI co-founder and CEO Pete Florence told me, referring to OpenAI’s groundbreaking large language model released in 2020. “If you take that model and prompt it to perform a new task, it will actually do it.”
Generalist AI appears focused on teaching robots an understanding of the physical world, sometimes called intuitive physical reasoning. Humans begin developing this ability early in life, and it may help AI models transfer knowledge from one situation to another. Several of the company’s demonstrations reminded me of the way children improvise and experiment when learning how to complete a task.
Researchers are often surprised by what these robots attempt. For example, when a banana was placed in front of one robot, it chose to use the banana to clean an object. The decision may seem simple, but physical intelligence remains a major challenge for machines. Understanding how babies learn so efficiently from the world around them could offer valuable insights for the future of robotics and artificial intelligence.
Source: www.wired.com