Teaching robots through positive reinforcement

There likely isn’t a robotics teacher institute in the world actively pursuing robotic learning. The field, after all, holds the key to unlocking a lot of potential for the industry. One of the things that makes it so remarkable is the myriad different approaches so many researchers are taking to unlock the secrets of helping robots essentially learn from scratch.

A new paper from Johns Hopkins University sporting the admittedly delightful name “Good Robot” explores the potential of learning through positive reinforcement. The title derives from an anecdote from author Andrew Hundt about teaching his dog to not chase after squirrels. I won’t go into that here — you can just watch this video instead:

But the core of the idea is to offer the robot some manner of incentive when it gets something correct, rather than a disincentive when it does something wrong. For robots, incentives come in the form of a scoring system — essentially a kind of gamification that rewards a number of points based on correctly executing a task.

The PhD candidate says the method was able to reduce the training time of a task significantly. “The robot wants the higher score,” Hundt said in a release tied to the research. “It quickly learns the right behavior to get the best reward. In fact, it used to take a month of practice for the robot to achieve 100% accuracy. We were able to do it in two days.”

The tasks are still quite elementary, including stacking bricks and navigating through a video game, but there’s hope that future robots will be able to work up to more complex and useful real-world tasks.