Using Fake Data To Help Robots Learn

Humans have the ability to learn on the fly, which makes us incredibly effective at adapting to our changing circumstances. Research from the University of Michigan explains how “fake” data could help robots do the same.

The researchers focused on helping robots learn how to work with various soft objects, such as fabrics and ropes, or even in cluttered environments. They believe that their new approach would cut the time required to learn from a few weeks to a few hours.

Learning the ropes

In various simulations, the training data was able to enhance the success rate of robots learning to loop a rope around an engine block by around 40%, while also doubling the success rate of a physical robot performing a similar task.

The task was chosen as it’s the kind of thing a robot mechanic would be required to perform. The researchers explain that learning to perform such a task would need vast quantities of data using current methods, which would make learning prohibitive.

“If the robot needs to play with the hose for a long time before being able to install it, that’s not going to work for many applications,” they explain.

Making generalizations

The new approach aims to help robots make the same kind of generalizations humans do, as these heuristics often allow us to transfer knowledge into new domains much faster than would be the case if we learned from scratch each time.

The researchers focused on three core qualities when developing their fabricated data, with it needing to be diverse, valid, and relevant.

“If you maximize the diversity of the data, it won’t be relevant enough. But if you maximize relevance, it won’t have enough diversity,” they explain. “Both are important.”

It’s as important that the data is valid. For instance, simulations that don’t practically work aren’t valid as the robot will be aware that such scenarios are practically impossible.

The researchers found that using just initial training data, the simulated robot was able to accurately hook the rope around the engine nearly half of the time, but after using the augmented data set, this rose to a 70% success rate.

Facebooktwitterredditpinterestlinkedinmail