Read our COVID-19 research and news.

Look out: Robots could soon teach each other new tricks

Someday soon, robot assistants will be a part of our everyday lives—but only if we can teach them new tasks without programming. If you have to learn to code, you might as well make the sandwich yourself. Now, a new system makes teaching robots almost as easy as teaching a child. And conveniently—or alarmingly, if you’re afraid of robot dominion—they can use this system to share their skills with each other.

There are two basic ways to train a robot. One is to program its movements, which requires time and coding expertise. The other is to demonstrate what you want by tugging on its limbs, moving digital representations of them, or doing the task yourself as an example for the robot to imitate. But delicate tasks sometimes require more precision than a person can demonstrate by hand—defusing a bomb is one good example. Now, with a system called C-LEARN, scientists have imbued a robot with a knowledge base of simple steps that it can intelligently apply when learning a new task.

“[C-LEARN] takes a very practical approach that works really well,” says Anca Dragan, a roboticist at the University of California, Berkeley, who was not involved in the research.

In this system, human users first help build the robot’s knowledge base. Researchers taught a two-armed robot called Optimus by clicking and dragging its limbs in a software program. They demonstrated movements, such as grasping the top of a cylinder or the side of a block. They performed each task seven times from different positions. The movement varied slightly each time, and the robot looked for patterns that it then integrated into its system. For example, if the grasper always ended up roughly parallel to the object, the robot would infer that parallelism was an important constraint to that process.

At this point, the robot is “like a 2-year-old baby that just knows how to reach for something and grasp it,” says Claudia Pérez D’Arpino, a computer scientist at the Massachusetts Institute of Technology in Cambridge and the leader of the study. With its knowledge base, the robot can learn new, multistep tasks with just a single demonstration. Users show robots the desired task with the C-LEARN software, and then approve or correct the robot’s attempt. It’s a one-and-done affair.

“Robots that can obey geometric constraints have been around for more than a decade,” says Maya Cakmak, a roboticist at the University of Washington in Seattle who was not involved in the work. “However, so far only experts have been able to make use of them.”

To test the system, the researchers taught Optimus four multistep tasks: to pick up a bottle and drop it in a bucket, to grab and lift a tray horizontally with both hands, to open a box with one hand and press a button inside it with the other, and to grasp a handle on a cube with one hand and pull a rod straight out of the cube with the other. For each task, Optimus received one demonstration and made 10 attempts. It succeeded 37 out of 40 times, researchers will report later this month at the IEEE International Conference on Robotics and Automation.

For an even tougher challenge, the researchers transferred Optimus’s knowledge base and its plans for the four tasks to a simulation of Atlas, a two-footed robot that has to keep its balance. Atlas managed to complete all four tasks. But when researchers deleted some of the transferred knowledge, such as the constraint of keeping certain movements parallel, it failed.

Such knowledge transfer would have practical application, D’Arpino says. “You can teach one robot to do something in a factory in Germany, and there’s no reason you can’t transfer that to a different robot in Canada.” Of course, of concern to those who have a dystopian view of the future is that robots teaching each other new skills over the internet would be a necessary first step toward world domination.

D’Arpino is now seeing whether people interacting with Optimus for the first time can teach it new tricks. The results so far are promising, though she’s not ready to discuss them in detail. Next, she hopes to teach robots the flexibility to adjust their learned skills on the fly.

One eventual goal is to teach the robots to disable bombs, a delicate task in which robots need to be directed quickly and with high precision. Other applications include finding people in a disaster, manufacturing electronics, and helping sick—or lazy—people with chores around the house. “There’s this promise of robots at home, but the reality is that now they can do nothing,” D’Arpino says. “What can a robot today do at your place, other than vacuum? It’s really hard.” She’s hoping to change that.