Open AI’s Robot Hand Learns Dexterity in Virtual Training

Open AI is a nonprofit AI research group that was founded in 2015 whose most recent research was backed by  Silicon Valley financier Sam Altman. The company is the brainchild of SpaceX founder Elon Musk.

On Monday, Open AI announced they had taught a robotic hand how to rotate a multicolored block with letters on it until the desired face of the block was pointing upwards. It sounds simple to us humans, doesn’t it?

But the interesting thing was how the robotic hand managed to gain that skill. First, it learned to do it in a virtual environment, and then that knowledge was transferred into the real world.

Physically training a robot takes months or even years and robotic hands, though they have been around for years, are notoriously difficult to program. So researchers have come up with a different solution, which involves distributing the physical training to multiple computers for the purpose of creating a software simulation. The simulation can do the training – without human help – in days or even hours, depending on complexity.

Credit: OpenAI / YouTube 

The system was called Dactyl and, to make it more reliable so that the hand would be able to cross the bridge between the simulation and actual, physical tasks, researchers placed random noise within the software simulation. This caused the robotic hand’s virtual world to be just messy enough so it would not be confused by the real world when it would start to operate within it.

The team is currently working on figuring out how to get the hand to perform more complicated tasks than turning a cube around, and we’re really curious as to what they have in store next.