Various researchers are finding new ways to teach their robots to do every task. For example, learning, teaching, playing, etc., make our lives easier, safer, and more entertaining.
Last month, a research company co-founded and chaired by Elon Musk- OpenAI had unveiled the world’s first Spam-detecting AI. This vision system was trained entirely in simulation, while the movement policy for grasping and removing the Spam is hard-coded. By using domain randomization, the company had trained its vision system. Now, the company has developed its latest version by deploying a new algorithm, one-shot imitation learning.
It allows a human to communicate how to do a new task by performing it in VR. Generally, one-shot imitation learning is commonly used to solve different tasks in isolation. The company’s primary goal is to develop a robot that should learn from very few demonstrations of any given task and instantly generalize to new situations of the same task without requiring task-specific engineering.
In actuality, developers used two different neural networks to develop their one-shot imitation learning system. 1. A vision network, and 2. An imitation network.
Primarily, the vision network takes an image from the robot’s camera and displays the states that show the positions of the objects. Engineers trained this vision network with the help of hundreds of thousands of simulated images with different perturbations of lighting, textures, and objects.
Later on, the imitation network observes the intent of a task; it observes a human demonstrating via a virtual simulation. It then outputs the task in the real-world setting. It is trained on dozens of different tasks with thousands of demonstrations for each task.
When the system starts imitating, it could do so after viewing the demonstration only once. The system was able to complete the task even if the starting parameters didn’t quite match up. The robot must learn how to infer the relevant portion of the task from the first demonstration to predict the action effectively.
It also performs attention over the locations of the different blocks, allowing it to imitate longer trajectories than it’s ever seen and stack blocks into a configuration with more blocks than any demonstration in its training data.
Unlike Musk’s other companies, OpenAI is a nonprofit organization. Its main goal is to support and guide AI innovation.