OpenAI Sets New Standard for Robotic Hand Dexterity


A robotic arm used in factories. / Photo by: 28 Panfilovs Men Studio via Wikimedia Commons


Nonprofit artificial intelligence company OpenAI has just set a new standard for robotic hand dexterity by teaching a robotic hand more elegant movements. This is according to The Verge.

The OpenAI team taught the robot hand to manipulate the six-sided cube printed with an uppercase letter in each side. After training the system, it can now do gestures close to what the human hand can achieve. These traits include finger pivoting, sliding, and finger gaiting. It means that the robotic hand can spin the block to show the letters of the certain side without dropping the block.

The trick the researchers used is reinforcement learning. During the simulation, the robotic hand experimented in different ways on how it can grasp the object and play with the block. OpenAI engineer Matthias Plappert said that in the beginning, the robot was just doing random things and even failing miserably. The researchers added random visual noise, changed the colors of the virtual hand, and alter the size of the toy cube or as to how heavy it was as part of the test. 

Over time, its experience made it more versatile to rotate the block in its robotic hand. The Verge published that the training was done to give the artificial intelligence a better understanding of what it is really like to manipulate the cube in the physical world.

The team called the system Dactyl. “Without this randomization, it would just drop the object all the time because it wasn’t used to it,” said Plappert, referring to the trial and error before the successful dexterity.