The saying ‘monkey see, monkey do’ has entire new meaning when robots start to learn by copying humans.
OpenAI, a non-profit artificial intelligence research company backed by Elon Musk, has announced a new algorithm focused on robots learning by imitation.
The algorithm allows humans to show a task to an AI by first demonstrating it in virtual reality (VR). According to OpenAI, this method is called one-time imitation learning.
OpenAI test this one-time imitation with a robotic arm that sorts and stacks blocks on a table top. The human does the stacking in VR, while the robotic arm imitates what the human does, however, in real life.
The robotic arm has two neural networks within its system, one called the vision network that ‘views’ and registers what actions are taking place in VR, and the other that neural network — called the imitation network — that guides the arm based on the information provided by the vision network.
The vision network does a lot before sending the information to the imitation network. First, within the network, it generates simulated images in a variety of colours and shapes of the table with the blocks. The simulated images generated help the vision network analyze what it “sees” with the camera feed.
The vision network then transfers the information to the imitation network that guides the robotic’s arm’s arm. Even though the blocks are set up are in a variety of colours and in different orders, the robot is still able to figure out the right order.
“Our robot has now learned to perform the task even though its movements have to be different than the ones in the demonstration,” said Josh Tobin a member of OpenAI, in a video demonstrating the new algorithm.
“With a single demonstration of a task, we can replicate it in a number of different initial conditions. Teaching the robot how to build a different block arrangement requires only a single additional demonstration.”
Source: Cornell University Library
Via: The Verge