Using AI, researchers devise a quick and accurate method for teaching robots. Complex skill

In Berkeley, California, Sergeyrevine’s robot AI researchers and learning love turned to the table where the 39 Genga block tower was completely piled up. Later, the white and white robot doubled the single limbs like a giraffe with a haterver, zooming towards the tower, and wielded a black leather whip.

Through something that seemed to be a casual viewer like a miracle of physics, the whip hit the accurate and appropriate place to send one block from the stack.

This task, known as “Jenga Whip”, is a hobby that dexterity and reflective people pursue it to do it. Now, thanks to a novel equipped with AI, it is learned by robots.

This training protocol teaches robots how to execute complex tasks like Genga Hipes with a 100 % success rate by learning from human demonstrations, feedback, and their own real world attempts.

In addition, robots are taught at an impressive speed, and you can learn how to build a computer motherboard completely within 1-2 hours and build shelves.

The robot learning field, which is supplied with fuel, is a method of teaching an unpredictable or complex machine activity in contrast to a single action, such as repeatedly picking up objects from a specific location of the conveyor belt. I tried to release it. To solve this puzzle, Levine’s lab has zero to what is called “enhanced learning.”

Jianlan Luo, a postdoc researcher, explained that in enhanced learning, robots learn from mistakes to try tasks in the real world, use feedback from the camera, and ultimately acquire their skills. When the team first announced a new software suite using this approach in early 2024, Luo says that others can use open source software to quickly reproduce success. I did it.

This fall, Levine, Ruo, Charles XU, Zheyuan Hu and Jeffrey Woo have released a technical report on the latest system that won Genga Hip. This new version and the improved version have been added in human intervention. The survey results are also published on the ARXIV Preplin Server.

By using a special mouse that controls the robot, humans can fix the robot course, and those modifications can be incorporated into the memory bank of the robot. Using an AI method called RenforceMent Learning, the robot analyzes the sum of all the trials (no support and support, success, failure) to execute the task.

Luo said, as robots have learned from experience, he needs to intervene more and more. “I had to babie a robot for the first 30 % or something, and I couldn’t actually pay much attention,” he said.

The lab put a robot system through a complicated task gauntlet beyond Genga Hip. The robot turned the egg over the pot. The object was handed from one arm to another arm. I assembled a motherboard, a car dashboard, and a timing belt. Researchers select these issues, and they are various, and Ruo’s words indicate “all kinds of uncertainty when running robot tasks in a complex real world.”

The timing belt task was outstanding in terms of difficulty. Every time the robot interacts with the timing belt, Imagine, who tries to operate the floppine neckless chain with two pegs, had to predict and respond to the change.

Genga Hipes configures different types of challenges. Because it contains physics that is difficult to model, training a robot using only a simulation is less efficient. The experience in the real world was important.

Researchers also tested the robot adaptability by stagering the accident. They force the gripper to open, and react to the shift status that may drop an object when the robot tries to install a microchip, move the motherboard, and encounter outside the lab environment. I trained to do.

By the end of the training, the robot can execute these tasks 100 % correctly. Researchers compared the results with the general “my actions” methods known as the behavior cloning trained with the same amount of demo data. Their new system has made robots faster and more accurate.

Luo said that these indicators were very important because the robot abilities were very high. Normal consumers and industrialists do not want to buy consistent robots. Luo emphasized that the “custom -made” manufacturing process that is often used in electronic devices, cars, and aerospace parts can benefit from robots that can ensure that various tasks are adapted.

When the robot first conquered the Genga Hiper Challenge, Ruo said, “It really shocked me.” “Genga tasks are very difficult for most humans. I tried it with a whip in my hand. I had a 0 % success rate.” He added that the robot was more likely to surpass humans because the robot did not have any tired muscles.

Levine Lab’s new learning system is part of a wider trend of Robotics Innovation. In the past two years, a larger field has moved dramatically, has been promoted to industry investment and AI, and engineers provide a turbo charging tool for analyzing performance data or image input that the robot may be observing. I will do it. Berkeley professors and researchers are part of this upwell in this innovation.

Levine has jointly established Robotics Company Physical Intelligence (Pi). This is currently being evaluated as $ 2 billion for progress in creating software functioning with various robots.

In 2018, Professor Kengoldberg and other Berkeley researchers formed Ambi Robotics. The company creates trained robots via AI simulation, grasps the parcels, classifies it into different containers, and is essential for e -commerce business.

Peter Abbeer, the director of the Berkeley Artificial Intelligence Research Lab, jointly created AI Robotics Startup Covariant, which was enlisted by models and braintrast last year. HOMAYOON KAZEROONI, a mechanical engineering professor, has established a public company EKSO Bionics, which sells robots “exiso sales” for those with limited mobility.

Regarding Luo’s research, he is excited to see where his team and other researchers can push it. One of the next steps is to train a system with a basic object operation function in advance, eliminate the need to learn from zero, and go straight to gain complex skills. He said. The lab also chose to create an open source for research so that other researchers could use it.

“The important goal of this project is to make this technology easy to access like an iPhone,” Luo says. “I believe that more people can use it, the greater we can have a greater impact.”