Robots Should Learn How to Improvise Before Entering Society
Source – sanvada.com
When a robot is inside of a lab, it is capable of doing many things, and this has much to do with its programming. But what happens when the AI meets the real world and the base programming is just not enough?
In order for robots to become useful in society at large, they cannot rely on specific tasks given to them via their programming. These machines must be able to learn and adapt to any situation, and as such, we believe creators should find a way to teach robots how to improvise.
By looking at factories around the world with robots at the helm, we can get a broad understanding of how these devices work. You see, robots are programmed to perform a specific task, but when it’s time to change to a different mission, the robot must first be reprogrammed.
That’s quite time-consuming; therefore, it would make more sense if robots could adapt to change without human intervention. That’s exactly what some researchers are hoping to accomplish in the distant future.
It’s not easy
Giving specific tasks to AI is quite easy for engineers and programmers, but designing a codebase that gives artificial intelligence the ability to adjust on the fly is not easy.
“It turns out those things are really hard,” said Cynthia Breazeal, a roboticist at the Massachusetts Institute of Technology’s Media Lab.
Not too long ago, MIT reported on how easy it is to trick an AI. The team created a 3D printed turtle, then proceeded to alter the lighting and coloring of the object.
When placed in front of an AI, the system was unable to tell the object was a turtle. Instead, the AI viewed it as a rifle, and right away it proves how easy it is to trick a robot.
In a world where everything is becoming automated and self-driving cars are potentially 40-years away from mass market adoption, a machine being easily manipulated is a major problem.
Tech companies are already working on creating more flexible robots
One of the companies at the forefront of change is no other than Google. The company’s DeepMind subsidiary uses a technique known as reinforcement learning to create software that has defeated humans’ multiple times.
The most notable was when Google’s AI defeated the best players at the Chinese game, Go.
It wasn’t perfect, but due to advanced machine learning, the artificial intelligence was able to learn and adapt its playstyle to become almost impossible to beat.
Real world doesn’t have a score
Robots working in the real world cannot rely on scores, according to Brown University roboticist Stefanie Tellex. In such a situation, creators must use something that is called “reward functions” to get things running as intended.
Reward functions is basically alerting the AI when it has acted accordingly. “The reward signal is so important to making these algorithms work,” Dr. Tellex added.
As it stands, for AI to perform and adapt in most situations, researchers might have to give reward them.