If you keep up with robotics, you may have heard about robots that can learn tasks. This idea raises a lot of thought provoking questions. What does it mean for a robot to learn? Can robots really learn? Why would we want robots to learn?
People become skilled at things through observation, learning, and practice. You’ve got to practice your instrument daily if you want to make it into Julliard. You need to hit the fields early and stay late if you plan on playing in the World Cup. Delivering the perfect speech requires hours of memorization and recitation. Even something as mundane as making an omelette requires practice if you’re going to get it right.
Robots, however, have it easy. They’re built and programmed for their specific task and are masters as soon as they’re deployed. Industrial machines don’t have to go through the motions for years and years before doing their task perfectly. They operate with constant precision, getting no better or worse, from load 1 to load 1 million.
So why would roboticists be working on robots that aren’t perfect from the get-go?
One of the advantages that robots have over people is the fact that they don’t need to learn how to do things, and that they are consistent and reliable in the things that they do. But along with that advantage is the disadvantage: robots can’t adapt to new needs and situations the way people do.
Researchers want robots to be able to adapt to changing environments, and handle unexpected, real-world situations. Robots are currently one-trick ponies, so while they are great at their specific task, they aren’t versatile, and changes in an environment can render a robot useless. Fabric, for example, has been robo-kryptonite with its non-rigid, billowy, unpredictable nature.
What researchers don’t want is for robots to learn at the same rate as humans. This has been one of the biggest challenges. Getting robots to learn is difficult enough, but the goal is to get robots to learn quickly, and to be able to adapt to changing circumstances as they happen.
Ways that researchers are currently trying to teach robots how to learn.
Good old-fashioned trial and error. Robots like Baxter are more versatile and capable of learning than traditional robots. These robots are equipped with programs that allow them to learn from experience.
Trial and error blitzkrieg. Google Research built an army of robotic arms and set them to the task of picking up different items with the aid of visual servoing. They picked up heavy objects, light objects, hard objects, soft objects, ones that were translucent, some that were big, and some that were small. These 14 robots, all grasping items simultaneously, produce large volumes of data that can be applied to the “learning” process.
So far, learning robots are pretty underwhelming. Picking up a pen or a pair of scissors just doesn’t seem that remarkable. That’s partly because we’re comparing them to a person’s ability to distinguish between a pair of scissors and a pen and pick either one up with ease, and with the awesome robots that can do incredible things in factories. Learning robots have to start somewhere, though, and the progress really has been impressive.
The fact that robots are able to learn and improve in any capacity is truly an accomplishment, and it seems very plausible that robots will be learning at an incredible rate in the future.