One of the biggest obstacles to the robot revolution is their inability to adapt to previously unseen situations. But that is about to change, thanks to a new approach that combines adaptive skills on the move to meet new challenges.
Recently, a team of joint researchers from Zhejiang University and the University of Edinburgh claims that they have developed a better way to teach robots how to walk and recover from falls on their own. In a recent article, the team explains how their dog-like robot, called Jueying, improvises new skills and adapts to unknown challenges in real-time. It uses the technique called multi-expert learning architecture (MELA) that learns to generate adaptive skills from a group of representative expert skills.
The team created the first trained software that could guide a virtual version of the robot. This software includes eight AI “virtual experts” that they have trained to master a specific skill. Among them, there are some virtual experts specializing in training on walking, balancing, and others who tend to teach robots how to stand up after a fall. Every time the digital robot successfully completed a task, a virtual expert group rewarded it with a virtual point.
After successfully training eight virtual experts in the latest AI coaching software suite, they developed a virtual network that acts as a type of head coach. This network manages the inputs of eight other algorithms, prioritizing one or more as needed in a given situation.
Eventually, they integrated their software into a prototype robot dog named Jueying. In the test, this dog was forced to fall to the ground, and suddenly, it was able to stand up and move normally in a short time.
Zhibin Li, a co-author of the report, told Wired that their team’s research goal was to create smarter machines, capable of rapidly combining flexible and adaptive skills, to tackle various tasks they had never seen before.
However, it may take a while before we see Jueying and Spot sparring for the best robot dog award. One of the challenges the team faced was reducing the amount of computing power required to simulate robot training. They need to do that to make it more useful for practical applications.