Can Robots Teach Each Other?

From bar tenders to personal assistants, robots will soon be accepted as a part of daily life. And according to experts, the final stages of the revolution hinges on the ability to teach robots new tasks, without the need for manual programming. Essentially, using robots to teach robots.

According to experts, there are two basic ways to train a robot. The first involves programming its movements, which calls for time and coding expertise. The second is to show the robot what’s required by actively moving its limbs, or physically demonstrating what’s required. However, delicate tasks often call for more precision and accuracy than a human can demonstrate first hand. For example, defusing a bomb.

Using software to train software

The innovative new C-LEARN programme solves both problems. Using intuitive software, it allows non-coders to teach robots complex tasks with a single demonstration. These skills can then be automatically transferred to other robots to perform similar actions.

So how does C-LEARN work? Working with a two-armed robot called Optimus, researchers built a knowledge base by clicking and dragging its limbs in the C-LEARN software program. Tasks were performed multiple times, with movements varying slightly for each sequence. As it learned the robot searched for patterns, which it then integrated into its system. For example, every time the robot lifted an object it noted that its grasper needed to be at a parallel angle. It then ‘learned’ that parallelism was a key constraint to the lifting process.

A simple train, correct, and approve model

Claudia Pérez D’Arpino, a computer scientist at the Massachusetts Institute of Technology led the study, and is confident that C-LEARN can help robots can master new, multistep tasks with just a single demonstration. Humans simply show robots the desired tasks, then correct and approve attempts.

By the end of its training, Optimus had mastered four multistep tasks. D’Arpino maintains that there are endless practical applications for C-LEARN, musing that “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.” Eventually, they could be smart enough to disable bombs, lead search and rescue parties, manufacture electronics, and even provide hospital care.

In the future, robots could hold their own in medical research labs like the Animal Replacement Centre of Excellence. For a closer look at the joint venture from the Queen Mary University of London (QMUL) and the Dr Hadwen Trust (DHT), ‘Human Models for Human Disease’ is a must read article.