Simulated Robot Demonstrates Unique Learning Capability

This post is also available in: עברית (Hebrew)

A simulated robot managed to carry on with a broken leg, proving that a continuous learning capability can help robots adapt to new circumstances. 

Most artificial intelligence (AI) relies on neural networks, which are algorithms based on the human brain. The main difference, however, is that the human brain continually keeps learning and adapting, whereas AI brains don’t. 

To further test this, researchers created nets with “Hebbian rules” — which are mathematical formulas that enable AI brains to continually learn, just like human brains. 

To test its theory, a research team partially removed the front left leg of two robots, to see how they would react in order to compensate for the missing limb. Both robots were a little shaky at first, but shortly thereafter the Hebbian bot carried on walking, whereas the regular AI brain bot flipped back and stopped. 

As the study pointed out, this is because only one of the robots has been taught to adapt to new circumstances. 

The study demonstrates how Hebbian learning could someday improve algorithms that are usually used for learning new languages, driving, or recognizing images. 

The study was published in NeurIPS Proceedings.