A Robot Maker Says Its Humanoid Did Chores in 30 Homes It Had Never Seen

Figure, a California company that builds human-shaped robots, says its robot picked up toys, folded towels and made beds in 30 Bay Area homes it had never been in, with no adjustment for any of them. It reports success on 56 percent of attempts, up from 9 percent without its new training method. The numbers are the company's own; no outside group has checked them yet.

Figure, a California company building humanoid robots, meaning robots shaped like a person, published results today of a test it says no one has run before. It sent its robot into 30 homes around the San Francisco Bay Area that the robot had never been in, and asked it to do three chores: pick up a dozen or so toys scattered around a living room and put them in a basket, fold towels, and make a bed. The company says the robot had no information about any of those homes ahead of time and was not adjusted for any of them. It went in and did the work on its own.

The robot is run by a program the company calls Helix 2.5. What is new is how it was taught. Figure has been assembling a very large collection of recordings of people doing everyday physical tasks, which it calls Index, and it used that collection to give the program a general feel for how rooms and objects behave before teaching it specific chores. In the company's own comparison, a version of the program trained on those human recordings succeeded on 56 percent of its attempts in unfamiliar homes; an otherwise identical version trained without them managed 9 percent. The company also says the new program needed only half as much chore-specific practice as its previous version, which had only been shown to work in the one house where it was trained.

The company's own account of why this matters: "A person can walk into an unfamiliar home and start working immediately... This is currently not true for robots." A robot that only works where it was trained is a demonstration. A robot that works in a stranger's living room is the beginning of something you could buy.

A few things to keep in view. These are the company's numbers; no outside group has checked them, and there is no scientific paper. A 56 percent success rate means the robot failed close to half the time, and the three tasks were the ones the company chose, not whatever a household happened to need. Figure itself says the result does not mean general-purpose home robots are solved, only that it now has early evidence that what its robot learns in one home carries over to the next. Whether that holds outside the Bay Area, and on messier jobs, is what the next tests will show.