Virtual Machine Learning Technique Could Help Robots Perform Better
Matthew Greenwood posted on January 24, 2019 |

An international team of researchers from Switzerland, Germany and the U.S. has found a way to boost the abilities of robots by using machine learning.

The team used machine learning to give a robot named ANYmal—a commercially available autonomous robot the size of a large dog that can walk and navigate its surroundings—greater speed and agility. ANYbotics, the company that makes it, wanted to give the robot the ability to learn through practice rather than through programming—an alternative that could save time and money and help enhance the robot’s abilities.

According to the researchers’ paper published in Science Robotics, robots with legs pose one of the biggest challenges in robotics: how do you teach them to walk? In particular, the dynamic and agile maneuvers of four-legged animals simply cannot be imitated or taught by existing methods devised by two-legged humans.

The team devised a promising alternative: allowing the robot to learn how to walk on its own by way of machine learning. Machine learning works by setting up goals for a robotic system and giving it the autonomy to try different ways to achieve that goal. As it learns, it sets new benchmarks and continues to improve. This process is repeated—sometimes thousands of times.

However, this kind of process is problematic in teaching a robot to walk because of the complexity of learning how to balance and the significant time investment. The team devised a way of reducing the time needed by creating a virtual version of the robot to do the learning on a computer desktop.

The virtual robot stumbled around learning to stand and maneuver for 11 hours before the team downloaded the results into the physical ANYmal. The virtual robot learned how to walk and balance itself about 1000 times faster than it would have in the real world.

It also learned how to perform better than the commercially available ANYmal. It ran 25 percent faster, was better able to prevent a human from kicking it over, and when it did fall, it got back up much quicker. In addition, it walked more precisely while using less torque and power.

The researchers recognize the need to refine the concept of virtual training for robots—but recognize that it shows great promise for teaching robots how to function faster and cheaper. “We see the results presented in this paper as a step toward comprehensive locomotion controllers for resilient and versatile legged robots.”

Learn more about the latest developments in robotics at Robot Laborer Could Do the Risky and Tiring Work of Human Laborers.

Recommended For You

Recommended For You