A new software system, called C-LEARN, allows humans to teach robots new tasks through simple demonstrations instead of through programming. Such a technique could make it much more practical to use robots for a wide range of tasks in the future. Generally, robots have had to learn tasks either by having the precise movements programmed through coding, a time and knowledge intensive task, or have the tasks demonstrated by tugging on its limbs or showing the robot a demonstration. However, certain tasks, such as defusing a bomb, require more precision than can be shown in a typical, by-hand demonstration.
C-LEARN teaches robots knowledge of simple steps they can apply in order to learn a new task.
According to UC Berkeley roboticist Anca Dragan, the software “takes a very practical approach that works really well.”
Researchers provided the knowledge base to a two-armed robot, named Optimus, by clicking and dragging its limbs using a software program. They demonstrated basic movements, like how to grab hold of objects of different shapes and sizes. Each task was performed seven times from different angles and positions. The movements varied each time, allowing Optimus to look for patterns and integrate them into the knowledge base. Claudia Pérez D’Arpino, the MIT computer scientitst who led the research, said that during this stage of the learning process, the robot is “like a 2-year-old baby that just knows how to reach for something and grasp it.”
After it has its knowledge base, the robot can quickly learn new tasks with only a single demonstration.
To test the C-LEARN system, Optimus was taught four multistep tasks, including picking up a bottle and placing it in a bucket, grabbing a tray and lifting it with both hands, opening a box with one hand while pressing a button inside the box with the other, and grasping a handle on a cube with one hand and pulling a rod from inside the cube with the other. Each task demonstrated to Optimus only once, and the robot was given 10 attempts for each task.
Optimus completed the tasks successfully in 37 out of 40 attempts.
This breakthrough would have broad practical applications, says D’Arpino.
“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.”